- The London opening keynote argues model capabilities are on an exponential while most organizations adopt AI linearly, and the developers in the room are the ones who close that gap and turn capability into products people use.
- Across three layers — models, the Claude platform, and Claude Code — Anthropic introduced new capabilities including managed-agents features (multi-agent orchestration, outcomes, dreaming) plus new self-hosted sandboxes and MCP tunnels, and Claude Code features like the agents view, routines, and CI autofix.
- The recurring theme is verification and async work: as agents run longer (from minutes to hours toward continuous), the default shifts from "I prompt Claude Code" to "I have Claude prompt Claude Code," with humans delegating work but keeping judgment.
Code with Claude London 2026: Opening Keynote
- Boris frames the moment as the distance between "I have an idea" and "it runs" collapsing again — the calculator feeling, except the calculator can now write a distributed system; Spotify merges over a thousand Claude Code PRs a month cutting migration time over 90%, and Binti cut foster-family licensing by 20 days.
- Capability is on an exponential (API volume up nearly 17x, average Claude Code developer running over 20 hours a week), but most organizations adopt linearly, creating a growing gap that developers close.
- Lisa notes Anthropic shipped eight frontier models in 12 months; she frames progress via "task horizon" (how long a model works before losing the thread), which moved from minutes a year ago to hours today, heading toward continuous, always-on, proactive agents.
- Developers should build for emerging capabilities, not just today's — design for the next version of Claude, keep making harder evals and prototypes, and treat model upgrades as a business opportunity; as models get smarter, generalized primitives like a file system and sandbox often beat heavy scaffolding.
- The advisor strategy splits execution from advising: a smaller model (Haiku/Sonnet) executes and calls a larger model (Opus) for advice, giving frontier quality at lower cost — Eve Legal reported frontier quality at five times lower cost.
- Claude Managed Agents is an agentic harness plus production infrastructure to ship roughly 10x faster; recent features include multi-agent orchestration, outcomes, and dreaming (introspecting past transcripts to self-improve), demoed via a fictional company "Counter" and its proactive Growthbot.
- Two new managed-agents features launched: self-hosted sandboxes (run execution on your own server, with first-class support for Daytona, Cloudflare, Vercel, and Modal) and MCP tunnels (securely reach internal MCP servers behind your firewall without exposing them publicly).
- Claude Code added interfaces and primitives for managing many agents ("multi-clauding"): the desktop app and a new CLI agents view, plus code review, mobile remote control, autofix, routines (run on a schedule, webhook, or API), and Claude Security; internal Anthropic use drove a 200% increase in PRs per engineer.
Task horizon — How long a model can work before losing the thread; a metric tracking progress from minutes to hours toward continuous work.
Scaffolding — The parts of an agent that aren't Claude — the loops, instructions, and tools — which can hold smarter models back.
Advisor strategy — Splitting execution (smaller model) from advising (larger model) so a cheaper model calls a frontier model for help when stuck.
Claude Managed Agents — An agentic harness paired with production-grade infrastructure to build and ship agents about 10x faster.
Multi-agent orchestration — Building and coordinating fleets of agents working together on a goal.
Outcomes — Specifying what success looks like so Claude iterates until it gets there.
Dreaming — A capability where Claude introspects on its previous transcripts to learn and self-improve.
Self-hosted sandbox — Running an agent's code execution on your own server/provider (Daytona, Cloudflare, Vercel, Modal) instead of Anthropic's.
MCP tunnel — A secure path letting managed agents reach internal MCP servers behind your firewall without exposing them publicly.
Routines — Async automations in Claude Code that run on a schedule, webhook, or API call — "a higher-order prompt."
CI autofix — A feature that babysits a PR to production, auto-fixing CI failures, review comments, and merge conflicts.
Morning London. Hello. This is the first time that we've taken Code with Claude outside of San Francisco, and I'm very excited to be here with everyone. Before we jump in, can I take a quick selfie? Is that okay? Ready? All right, I think we got it.
I want to start today by telling you how I originally got into coding. I actually learned on a scientific calculator, and I found out that I could program my calculator to help me pass math tests and do well on homework. I was using a TI-83 calculator — I think in the UK you use Casios for this. I found out you could write little programs in TI-BASIC, the BASIC dialect for TI calculators, and reference them when I couldn't remember what to do. It worked beautifully. I got great scores, then I taught my classmates how to program their calculators and they also got great scores. At some point I even published a guide on the internet for how to program scientific calculators — I was 13 when I wrote this. That was my introduction to software. It was very practical, and as a kid it just felt like magic. Around the same age, I figured out HTML — not to build a startup, but to make my eBay listings look better than everyone else's so I could sell Pokémon cards. There were tables, blinking text, the whole thing. And it worked. I sold three legendary birds for 99 cents each, all holographic, and in hindsight I wish I'd held on to those.
There was a whole generation of us like that. We didn't learn to program from a textbook — we learned from tinkering. And I still remember the giddy feeling I got when one of my calculator programs ran. You made the thing and it did what you wanted. Then somewhere along the way, programming got complicated: compilers, type checkers, build systems, package managers, 12 config files before you could write a single line of code. The distance between "I have an idea" and "it runs" kept getting longer. What's happening now is that distance is collapsing again. You describe a problem and the program shows up. It's the calculator feeling, except the calculator can write a distributed system. Now I get to watch people experience the same feeling of bringing ideas to life, on a scale none of us imagined even a year ago.
It's gratifying to talk to people experiencing this every day, from indie devs to our biggest customers taking on significant challenges. For instance, Spotify uses Claude Code to migrate thousands of repos. The team led by Nicholas Gustafsson built a background agent on Claude that reads a migration described in plain English and runs it across a fleet of agents, opening PRs. It's now merging over a thousand PRs a month into production and cutting migration time by over 90%. That's real engineering hours back. And sometimes speed isn't just about efficiency — it's about delivering on a mission. Felicia Curcuru is the co-founder and CEO of Binti. Her software runs the systems caseworkers use to place kids in foster care — the paperwork, home visits, licensing. This year her team used the Claude API to give caseworkers back hours they used to spend on paperwork, and took 20 days off the process of licensing a foster family. Twenty days. That's not just an efficiency metric — that's a kid connecting with a family.
The industry is moving at incredible speed, and that makes sense when you look at the exponential and how fast these models are getting better. A couple of years ago, the frontier was good enough that Claude could draft a pretty good git commit. A year ago we were on stage at our first Code with Claude event in San Francisco; Opus 4 was the headline, and the idea that Claude could run for a few minutes and build an entire feature without a human checking in was crazy. Six months ago, agents started running end-to-end overnight and I started waking up to finished work. Last month, Mythos read the entire OpenBSD source tree and found a 27-year-old vulnerability that survived every human reviewer, fuzzer, and static analyzer for almost three decades. The jumps keep getting bigger and the intervals keep getting shorter. But even though model capabilities are improving on an exponential, most organizations are still adopting AI on a linear path. That means there's a growing gap between what AI can do and what it's actually doing for people. Closing that gap — translating model capability into something people can actually use — is what you all as developers do. API volume is up nearly 17x on the Claude platform, and on Claude Code the average developer is now spending over 20 hours a week running Claude.
This morning we're going to show you what that looks like. First, Lisa will talk about our foundation — the model layer. On the Claude platform, Angela and Caitlyn will walk you through how you can securely build and deploy agents at scale with Claude Managed Agents. And on Claude Code, Cat and I will walk you through the latest features like the new agents view and new primitives like routines, to let Claude prompt itself even when you're away from your computer. It all comes back to you and what you'll build, because most people are never going to call the Claude API or run Claude in a terminal — they'll experience AI through something one of you built. A designer exploring directions with Canva, a lawyer getting a brief out faster with Lora, or a developer using any of the world's best coding agents. Thanks for partnering with us to shape what AI looks like for the world. Up next, Lisa from the research PM team.
Hi. I joined Anthropic in 2023 and have been part of every model launch since Claude 3. That means I've helped bring 17 different versions of Claude to users and developers. Claude has come a long way. Opus 3 launched just over two years ago and was our first model proficient at writing long-form code. Sonnet 3.6 was the first that could use a computer safely. Sonnet 3.7 was our first that would think before answering. Last year at this event I introduced Opus 4 — which, we didn't quite know at the time, was our first model that could write complex Excel files and PowerPoint documents. And now, with Opus 4.7 and the Mythos preview, Claude can own outcomes end-to-end and apply judgment to complete tasks with high ambiguity. We haven't slowed down — we've shipped eight frontier models in the past 12 months, each building on the last.
We talk a lot about the exponential at Anthropic. We believe as model intelligence increases, the value of use cases increases exponentially — agentic coding is far more impactful than code autocomplete. Incremental intelligence creates new markets and grows the pie. What's really exciting is when Claude can do something it couldn't before: tool use, computer use, adaptive thinking, agentic loops that hold a plan over hundreds of steps, long context windows. We pioneered these areas, so when you build on Claude you're building on the model family that shipped these capabilities first and has had the longest to make them reliable. These foundations now support a diversity of tasks across domains — Claude can iterate on marketing presentations, redline legal documents, build financial forecasts, even analyze genomic sequencing data. AMP moved their smart mode to Opus 4.7 after it outperformed on their benchmarks and simplified their tooling. Rakuten found it resolved three times more production engineering tasks than the previous model and saw it catching its own logical faults during planning. We launched Claude Design from Anthropic's labs, powered by Opus 4.7, and customers are building production interfaces in a single sitting — we've heard Opus 4.7 has a taste for visual design. Users view Claude as a valued thought partner since it's willing to challenge their beliefs and push back when they're wrong. But models are still imperfect — Claude has verbal tics, can be stumped by viral common-sense questions, and sometimes does more than you asked.
Here's what to look forward to: higher judgment and code taste, so models you can trust with complex autonomous engineering; context windows that feel effectively infinite as models take on longer work without losing intent; and multi-agent coordination powering teams of agents on goals too big for any one agent. One metric I use is task horizon — how long can a model work before losing the thread. Last year, models could reliably work for minutes; today most users have agents that run for hours. We expect future generations of Claude to run continuously — proactive, always-on agents that know what to do without being told, responsible for high-level goals requiring judgment and collaboration. Instead of asking Claude to write a project update, you'd ask Claude to keep the project on track this week. Instead of asking for a financial forecast, Claude would own and update the forecast to keep it accurate over time.
So what should developers do? Since the exponential keeps moving, future Claude models will be more capable than Opus 4.7 and even Mythos preview. You need to build for emerging capabilities, not just what works today — design for the next version of Claude. The developers who win are the ones whose architecture is ready to absorb the next big jump. Scaffolding is what we call the parts of the agent that aren't Claude — the loops, instructions, tools. As models get smarter, scaffolding that used to help can hold Claude back; more intelligent models can often get further with generalized primitives like a file system and sandbox computing environment. You'll also need to keep making harder evals and product prototypes — they're how you notice the exponential moving underneath you. When a task that used to fail starts passing, that's your sign to ship something you couldn't before. And the teams getting the most out of Claude treat model upgrades as a business opportunity — make upgrades easy by automating your evaluations and testing, and test models hands-on. Now, Caitlyn and Angela will show you how the Claude platform makes this reality come to life.
Model capabilities are on this exponential, but most businesses are still on the linear. What's preventing businesses from snapping to the exponential boils down to two problems. The first is that getting the right outcomes can be too difficult — building an agent that does exactly what you need requires prompt optimization, tool construction, harness engineering, so much to get it just right. The second is you need to ship fast but scalably at the same time; it's easy to build prototypes but hard to scale in production. We built the Claude platform to give you everything you need: API primitives tuned to Claude models, infrastructure to build and scale agentic systems, and controls to operate those systems.
One of the most common problems we hear is that businesses need frontier-level intelligence but at lower cost. One way we're solving this is the advisor strategy. It's easy to implement — you just update your tools array on the messages API. Behind the scenes we split execution from advising: in execution you can use a smaller model to make it cheaper, but when that small model needs help, it can reach out to a larger model for advice. In practice you can have a Haiku or Sonnet-class model execute and use Opus as an adviser. We ran this with Sonnet as executor and Opus as adviser, and Sonnet performed way better than Sonnet alone — and even more cheaply, because Opus advised it to get its work done better. Eve Legal used the advisor strategy and got frontier model quality at five times lower cost. It's perfect when you want to put your best foot forward but stay conscious of cost, and great for extremely high-volume agentic workloads where you have to watch ROI.
Now, speed and scale. We recently introduced Claude Managed Agents — an agentic harness paired with production-grade infrastructure that lets teams ship literally 10 times faster, building a production-grade agent in days, not months. One favorite example is Asana, which built AI teammates so humans can collaborate directly with agents within Asana projects, delegating tasks. They built on Claude Managed Agents for speed and scale. Recently we upgraded it with multi-agent orchestration (build fleets of agents), outcomes (specify what success looks like and Claude iterates to get it done), and dreaming (Claude introspects on its previous transcripts to learn and self-improve). Today we're introducing two more: self-hosted sandboxes, so Claude can execute work on your own server, and MCP tunnels, so people who build internal agents can securely access their internal MCP servers. Claude Managed Agents spawn sandboxes when it's time to execute work like editing files; you can use ours or, starting today, your own — with first-class support for Daytona, Cloudflare, Vercel, and Modal sandboxes, with easy integration. With MCP tunnels, your MCP servers can be hosted behind your firewall on your private network so Claude Managed Agents can access them without exposing them over the public internet.
Caitlyn and I have been working with a fictional company called Counter, which offers software for small businesses to build a digital storefront. Counter wants to become an AI-native business, and one of its most common workflows is constantly running growth experiments to optimize merchant onboarding. So here we are in Counter's Slack workspace, with my colleague Gabe, wearing our cool Counter hats. We're excited to see how our latest growth experiment on the onboarding flow performed. Our new Growthbot came into the Slack channel and told us there's a clear winner: the simpler version of the onboarding flow is outperforming the older, longer one where merchants dropped off before finishing signup. The agent proactively tells us it's calling the experiment and getting started on cleaning up the old variant.
Let's look at how the agent is set up, using the Claude API CLI and the Claude developer console. First the agent configuration — system prompt, skills, tools. Then the MCP servers: Slack (the bot talked to us in Slack), a data warehouse so the bot can read experiment data, and a feature flags MCP server so the bot can call experiments or start new ones. Look at the URLs: Slack is a regular public MCP URL, but the data warehouse and feature-flags services are behind tunnel.anthropic.com — both MCP servers are behind Counter's firewall, only accessible via the tunnel, so Counter keeps its MCP servers on its private network. In the console, Counter set up a gateway in their own private network and established a secure connection to Anthropic, so any agent Caitlyn builds that uses the MCP tunnel can securely access internal MCP servers. You can create these tunnels directly in the console with a couple of configuration options. Counter's next requirement is that when the agent writes and executes code, it does so on a self-hosted sandbox, not public infrastructure. So for our managed agents we set up the environment to use a self-hosted sandbox — when the agent needs to execute work, it puts a work item in a queue, and here Counter uses Vercel's self-hosted sandbox, which picks up the item and spins up a sandbox in Counter's Vercel account to execute the work.
Managed agents have rich observability in the console. Let's check the session that produced that Slack thread. The agent runs a bunch of tools; we can confirm the things we cared about. There's a "set flag winner" tool call, meaning the agent worked through the MCP tunnel to call the feature-flag service. And when the agent writes the code to clean up the old variant, it checks out a new git branch and does this work on our Vercel self-hosted sandbox. Back in Slack, Growthbot is very proactive: it figured out it should default the simplified variant and put together a PR to do the defaulting and clean it up, executed on Counter's servers via the self-hosted sandbox. It even gave us a screenshot of the messy old variant versus the cleaner new one. It's also collaborative — when Caitlyn asked "What next?", Growthbot detected another opportunity to solve a 46% drop-off it found in the onboarding flow, calculated because it queried the data warehouse securely through an MCP tunnel, and asked if we wanted it to get started. Thank you, Claude. To wrap up: today we introduced two big upgrades to Claude Managed Agents — self-hosted sandboxes and MCP tunnels — making it easier than ever to become an AI-native business and build agents for all your workflows. Now Cat and Boris will talk about how Claude Code is making it even more fun to ship as a developer.
Angela and Caitlyn just showed how the Claude platform bridges the gap between what models can do and what businesses ship. We see a related challenge on Claude Code: bridging the gap between model capabilities and how every developer can benefit from them. First, thank you to all the builders in the room and watching online — thank you for trusting us and your production codebase back when Sonnet 3.7 was our frontier model and our product was rough around the edges. Our mission is to help every builder close the gap between having a great idea and having a shipped product. We do that by building tools that elicit the frontier intelligence from our models and making them accessible to every builder. We don't think of ourselves as having a finished roadmap — more like mountaineers climbing alongside you in terrain none of us has mapped, figuring things out as we go.
Last year I read every single edit Claude Code tried to make and every permission prompt, giving super-detailed guidance. Now most of us run in auto mode to delegate permissions, and we check in after Claude has tested, verified its changes, and has a PR ready. Claude Code started in the CLI, and the terminal is still the interface for power users who want minimal text and the most control. Then we added the IDE — the same agent, but you can follow along with code changes. Many of you now juggle multiple Claude Code instances, which we affectionately call multi-clauding. We added two new interfaces to manage more agents. One I use frequently is Claude Code on desktop — a full-screen graphical interface with built-in previews, a sidebar control plane, and the ability to render images and rich outputs. Desktop is a single view across local and cloud sessions, with visual indicators of which agents are running, blocked, or need input. Next is our newest surface, the Claude agents view in the CLI for people who prefer to stay in the terminal — see what's running, waiting, and done at a glance, reply inline to unblock, and jump in and out of sessions without losing your place. The VS Code extension and desktop app are both built on the Claude Agent SDK, the same one many of you build on. Many enterprises have adopted Claude Code wall-to-wall — at Anthropic it's driven a 200% increase in PRs per engineer even as our org has scaled substantially.
Here's some feedback we heard and what we built. You wanted to spend less time on code review, so we shipped a code review product that deploys a team of agents to traverse all your code changes and catch critical bugs — thousands of companies use it daily, including every internal Anthropic team. You wanted to code on the go, so we launched remote control in Claude Code on iOS and Android so you can fire off a task anywhere — go to a park, touch grass, and still get tasks done. You spend a lot of time babysitting PRs, fixing flaky CI tests, addressing review comments, and resolving merge conflicts, so we added autofix that listens for these events and proactively fixes each so your PR is always green. You wanted to run Claude Code on new tickets and customer bug reports, so we built routines: configure once and Claude Code runs on a schedule or in response to a webhook or API request. And you're landing so much code that security teams struggle to keep up, so we built Claude Security, which scans your codebase overnight, flags vulnerabilities ranked by severity, and lets you kick off Claude Code to address each. These primitives compose together. Companies have adopted Claude Code at the scale of entire engineering orgs: Shopify uses it across engineering and non-engineers (PMs, designers, data scientists) and brings it into their platform — Andrew McNamara, director of applied AI, says the speed is just crazy. Mercado Libre, Latin America's most popular e-commerce site with over 100 million buyers and 23,000 engineers, runs on Claude Code — it's reviewed more than 500,000 PRs with human oversight and modernized more than 9,000 apps, and Oscar (who leads technology) aims for 90% autonomous coding in a fully agent-driven PR loop by Q3. The detail I love most is that managers and VPs who haven't committed code in years are shipping again. Now, please welcome the creator of Claude Code, Boris Cherny.
I just realized we're matching. Thanks, Cat. Everything we're showing still feels magical to me — using Claude Code I get the same feelings I had programming my first calculator. Inside Anthropic, all day we're trading screenshots of things engineers are doing in the wild with Claude Code. For this demo, imagine we're an engineer at Acme Pay, a payment infrastructure company. We start in the Claude Code desktop app on one task: Claude is adding refunds to Acme's merchant dashboard — full implementation, idempotency so a duplicate webhook doesn't double-refund, multi-currency handling across regions, and audit logging for compliance. It writes the implementation and verifies its own work: it pulls up the merchant dashboard in the browser, triggers a refund, and the modal closes before the success toast appears — a real edge case. Claude sees the failure, traces it to a race condition in the optimistic update, fixes it, and verifies it works in the browser before calling the task done.
Now zoom out. Claude Code isn't running just one session — it's one of many running and managed in parallel. In the desktop app you can see all your Claude Code sessions: which are running, which need input, which have PRs merged and closed. Synchronous coding is now just a slice of what's happening, and going forward more code will be written async. That's why we keep talking about verification — if Claude can check its work, you can let it run while you work on something else and come back to a fully working result. A lot of my code these days is written by routines — I'm not doing the prompting, I create a routine that does the prompting. For engineers, think of it like a higher-order function: routines are a higher-order prompt. Back to the refund session — a teammate found a GitHub issue overnight, a routine watching the repo picked it up async and kicked off the work. With routines, developers set up async automations and wake up to PRs ready to merge. Routines can run on a schedule, be kicked off by webhooks or arbitrary API calls, and run locally or on remote cloud compute.
One more desktop feature: CI autofix. It watches the PR the last session opened and babysits it to production — auto-fixing comments from code review and security review, auto-fixing CI, and auto-rebasing on merge conflicts. Look what just happened: CI flicked on a network timeout, the routine woke up, diagnosed it as a known infra issue, and retried the job — and in the Claude Code codebase we tell it to fix the root cause instead of just retrying. CI is green, the engineer who owns the PR never sees the red X, and that work is off their plate. That's the shift: the default isn't "I'm going to prompt Claude Code," it's "I'm going to have Claude prompt Claude Code." Everything you just saw is available today, including routines and all the latest desktop updates. We hope these features let you close the gap between your ideas and shipping products.
That's what every talk today was pointing at — Lisa's capability curve, Angela and Caitlyn's agents that run on infrastructure you control, and what Cat and I just showed. These are three layers of one story: the capability is already here, and the remaining gap is how fast we put it to work. Spend the rest of today exploring these layers — research talks if you're evaluating models, Claude platform sessions if you're building for your users, and Claude Code workshops if you want more ways to bring Claude into your day-to-day workflow. Dive in, go deep, and start building with us. Thank you.
TL;DR
- The London opening keynote argues model capabilities are on an exponential while most organizations adopt AI linearly, and the developers in the room are the ones who close that gap and turn capability into products people use.
- Across three layers — models, the Claude platform, and Claude Code — Anthropic introduced new capabilities including managed-agents features (multi-agent orchestration, outcomes, dreaming) plus new self-hosted sandboxes and MCP tunnels, and Claude Code features like the agents view, routines, and CI autofix.
- The recurring theme is verification and async work: as agents run longer (from minutes to hours toward continuous), the default shifts from "I prompt Claude Code" to "I have Claude prompt Claude Code," with humans delegating work but keeping judgment.
Takeaways
- Boris frames the moment as the distance between "I have an idea" and "it runs" collapsing again — the calculator feeling, except the calculator can now write a distributed system; Spotify merges over a thousand Claude Code PRs a month cutting migration time over 90%, and Binti cut foster-family licensing by 20 days.
- Capability is on an exponential (API volume up nearly 17x, average Claude Code developer running over 20 hours a week), but most organizations adopt linearly, creating a growing gap that developers close.
- Lisa notes Anthropic shipped eight frontier models in 12 months; she frames progress via "task horizon" (how long a model works before losing the thread), which moved from minutes a year ago to hours today, heading toward continuous, always-on, proactive agents.
- Developers should build for emerging capabilities, not just today's — design for the next version of Claude, keep making harder evals and prototypes, and treat model upgrades as a business opportunity; as models get smarter, generalized primitives like a file system and sandbox often beat heavy scaffolding.
- The advisor strategy splits execution from advising: a smaller model (Haiku/Sonnet) executes and calls a larger model (Opus) for advice, giving frontier quality at lower cost — Eve Legal reported frontier quality at five times lower cost.
- Claude Managed Agents is an agentic harness plus production infrastructure to ship roughly 10x faster; recent features include multi-agent orchestration, outcomes, and dreaming (introspecting past transcripts to self-improve), demoed via a fictional company "Counter" and its proactive Growthbot.
- Two new managed-agents features launched: self-hosted sandboxes (run execution on your own server, with first-class support for Daytona, Cloudflare, Vercel, and Modal) and MCP tunnels (securely reach internal MCP servers behind your firewall without exposing them publicly).
- Claude Code added interfaces and primitives for managing many agents ("multi-clauding"): the desktop app and a new CLI agents view, plus code review, mobile remote control, autofix, routines (run on a schedule, webhook, or API), and Claude Security; internal Anthropic use drove a 200% increase in PRs per engineer.
Vocabulary
Task horizon — How long a model can work before losing the thread; a metric tracking progress from minutes to hours toward continuous work.
Scaffolding — The parts of an agent that aren't Claude — the loops, instructions, and tools — which can hold smarter models back.
Advisor strategy — Splitting execution (smaller model) from advising (larger model) so a cheaper model calls a frontier model for help when stuck.
Claude Managed Agents — An agentic harness paired with production-grade infrastructure to build and ship agents about 10x faster.
Multi-agent orchestration — Building and coordinating fleets of agents working together on a goal.
Outcomes — Specifying what success looks like so Claude iterates until it gets there.
Dreaming — A capability where Claude introspects on its previous transcripts to learn and self-improve.
Self-hosted sandbox — Running an agent's code execution on your own server/provider (Daytona, Cloudflare, Vercel, Modal) instead of Anthropic's.
MCP tunnel — A secure path letting managed agents reach internal MCP servers behind your firewall without exposing them publicly.
Routines — Async automations in Claude Code that run on a schedule, webhook, or API call — "a higher-order prompt."
CI autofix — A feature that babysits a PR to production, auto-fixing CI failures, review comments, and merge conflicts.
Transcript
Morning London. Hello. This is the first time that we've taken Code with Claude outside of San Francisco, and I'm very excited to be here with everyone. Before we jump in, can I take a quick selfie? Is that okay? Ready? All right, I think we got it.
I want to start today by telling you how I originally got into coding. I actually learned on a scientific calculator, and I found out that I could program my calculator to help me pass math tests and do well on homework. I was using a TI-83 calculator — I think in the UK you use Casios for this. I found out you could write little programs in TI-BASIC, the BASIC dialect for TI calculators, and reference them when I couldn't remember what to do. It worked beautifully. I got great scores, then I taught my classmates how to program their calculators and they also got great scores. At some point I even published a guide on the internet for how to program scientific calculators — I was 13 when I wrote this. That was my introduction to software. It was very practical, and as a kid it just felt like magic. Around the same age, I figured out HTML — not to build a startup, but to make my eBay listings look better than everyone else's so I could sell Pokémon cards. There were tables, blinking text, the whole thing. And it worked. I sold three legendary birds for 99 cents each, all holographic, and in hindsight I wish I'd held on to those.
There was a whole generation of us like that. We didn't learn to program from a textbook — we learned from tinkering. And I still remember the giddy feeling I got when one of my calculator programs ran. You made the thing and it did what you wanted. Then somewhere along the way, programming got complicated: compilers, type checkers, build systems, package managers, 12 config files before you could write a single line of code. The distance between "I have an idea" and "it runs" kept getting longer. What's happening now is that distance is collapsing again. You describe a problem and the program shows up. It's the calculator feeling, except the calculator can write a distributed system. Now I get to watch people experience the same feeling of bringing ideas to life, on a scale none of us imagined even a year ago.
It's gratifying to talk to people experiencing this every day, from indie devs to our biggest customers taking on significant challenges. For instance, Spotify uses Claude Code to migrate thousands of repos. The team led by Nicholas Gustafsson built a background agent on Claude that reads a migration described in plain English and runs it across a fleet of agents, opening PRs. It's now merging over a thousand PRs a month into production and cutting migration time by over 90%. That's real engineering hours back. And sometimes speed isn't just about efficiency — it's about delivering on a mission. Felicia Curcuru is the co-founder and CEO of Binti. Her software runs the systems caseworkers use to place kids in foster care — the paperwork, home visits, licensing. This year her team used the Claude API to give caseworkers back hours they used to spend on paperwork, and took 20 days off the process of licensing a foster family. Twenty days. That's not just an efficiency metric — that's a kid connecting with a family.
The industry is moving at incredible speed, and that makes sense when you look at the exponential and how fast these models are getting better. A couple of years ago, the frontier was good enough that Claude could draft a pretty good git commit. A year ago we were on stage at our first Code with Claude event in San Francisco; Opus 4 was the headline, and the idea that Claude could run for a few minutes and build an entire feature without a human checking in was crazy. Six months ago, agents started running end-to-end overnight and I started waking up to finished work. Last month, Mythos read the entire OpenBSD source tree and found a 27-year-old vulnerability that survived every human reviewer, fuzzer, and static analyzer for almost three decades. The jumps keep getting bigger and the intervals keep getting shorter. But even though model capabilities are improving on an exponential, most organizations are still adopting AI on a linear path. That means there's a growing gap between what AI can do and what it's actually doing for people. Closing that gap — translating model capability into something people can actually use — is what you all as developers do. API volume is up nearly 17x on the Claude platform, and on Claude Code the average developer is now spending over 20 hours a week running Claude.
This morning we're going to show you what that looks like. First, Lisa will talk about our foundation — the model layer. On the Claude platform, Angela and Caitlyn will walk you through how you can securely build and deploy agents at scale with Claude Managed Agents. And on Claude Code, Cat and I will walk you through the latest features like the new agents view and new primitives like routines, to let Claude prompt itself even when you're away from your computer. It all comes back to you and what you'll build, because most people are never going to call the Claude API or run Claude in a terminal — they'll experience AI through something one of you built. A designer exploring directions with Canva, a lawyer getting a brief out faster with Lora, or a developer using any of the world's best coding agents. Thanks for partnering with us to shape what AI looks like for the world. Up next, Lisa from the research PM team.
Hi. I joined Anthropic in 2023 and have been part of every model launch since Claude 3. That means I've helped bring 17 different versions of Claude to users and developers. Claude has come a long way. Opus 3 launched just over two years ago and was our first model proficient at writing long-form code. Sonnet 3.6 was the first that could use a computer safely. Sonnet 3.7 was our first that would think before answering. Last year at this event I introduced Opus 4 — which, we didn't quite know at the time, was our first model that could write complex Excel files and PowerPoint documents. And now, with Opus 4.7 and the Mythos preview, Claude can own outcomes end-to-end and apply judgment to complete tasks with high ambiguity. We haven't slowed down — we've shipped eight frontier models in the past 12 months, each building on the last.
We talk a lot about the exponential at Anthropic. We believe as model intelligence increases, the value of use cases increases exponentially — agentic coding is far more impactful than code autocomplete. Incremental intelligence creates new markets and grows the pie. What's really exciting is when Claude can do something it couldn't before: tool use, computer use, adaptive thinking, agentic loops that hold a plan over hundreds of steps, long context windows. We pioneered these areas, so when you build on Claude you're building on the model family that shipped these capabilities first and has had the longest to make them reliable. These foundations now support a diversity of tasks across domains — Claude can iterate on marketing presentations, redline legal documents, build financial forecasts, even analyze genomic sequencing data. AMP moved their smart mode to Opus 4.7 after it outperformed on their benchmarks and simplified their tooling. Rakuten found it resolved three times more production engineering tasks than the previous model and saw it catching its own logical faults during planning. We launched Claude Design from Anthropic's labs, powered by Opus 4.7, and customers are building production interfaces in a single sitting — we've heard Opus 4.7 has a taste for visual design. Users view Claude as a valued thought partner since it's willing to challenge their beliefs and push back when they're wrong. But models are still imperfect — Claude has verbal tics, can be stumped by viral common-sense questions, and sometimes does more than you asked.
Here's what to look forward to: higher judgment and code taste, so models you can trust with complex autonomous engineering; context windows that feel effectively infinite as models take on longer work without losing intent; and multi-agent coordination powering teams of agents on goals too big for any one agent. One metric I use is task horizon — how long can a model work before losing the thread. Last year, models could reliably work for minutes; today most users have agents that run for hours. We expect future generations of Claude to run continuously — proactive, always-on agents that know what to do without being told, responsible for high-level goals requiring judgment and collaboration. Instead of asking Claude to write a project update, you'd ask Claude to keep the project on track this week. Instead of asking for a financial forecast, Claude would own and update the forecast to keep it accurate over time.
So what should developers do? Since the exponential keeps moving, future Claude models will be more capable than Opus 4.7 and even Mythos preview. You need to build for emerging capabilities, not just what works today — design for the next version of Claude. The developers who win are the ones whose architecture is ready to absorb the next big jump. Scaffolding is what we call the parts of the agent that aren't Claude — the loops, instructions, tools. As models get smarter, scaffolding that used to help can hold Claude back; more intelligent models can often get further with generalized primitives like a file system and sandbox computing environment. You'll also need to keep making harder evals and product prototypes — they're how you notice the exponential moving underneath you. When a task that used to fail starts passing, that's your sign to ship something you couldn't before. And the teams getting the most out of Claude treat model upgrades as a business opportunity — make upgrades easy by automating your evaluations and testing, and test models hands-on. Now, Caitlyn and Angela will show you how the Claude platform makes this reality come to life.
Model capabilities are on this exponential, but most businesses are still on the linear. What's preventing businesses from snapping to the exponential boils down to two problems. The first is that getting the right outcomes can be too difficult — building an agent that does exactly what you need requires prompt optimization, tool construction, harness engineering, so much to get it just right. The second is you need to ship fast but scalably at the same time; it's easy to build prototypes but hard to scale in production. We built the Claude platform to give you everything you need: API primitives tuned to Claude models, infrastructure to build and scale agentic systems, and controls to operate those systems.
One of the most common problems we hear is that businesses need frontier-level intelligence but at lower cost. One way we're solving this is the advisor strategy. It's easy to implement — you just update your tools array on the messages API. Behind the scenes we split execution from advising: in execution you can use a smaller model to make it cheaper, but when that small model needs help, it can reach out to a larger model for advice. In practice you can have a Haiku or Sonnet-class model execute and use Opus as an adviser. We ran this with Sonnet as executor and Opus as adviser, and Sonnet performed way better than Sonnet alone — and even more cheaply, because Opus advised it to get its work done better. Eve Legal used the advisor strategy and got frontier model quality at five times lower cost. It's perfect when you want to put your best foot forward but stay conscious of cost, and great for extremely high-volume agentic workloads where you have to watch ROI.
Now, speed and scale. We recently introduced Claude Managed Agents — an agentic harness paired with production-grade infrastructure that lets teams ship literally 10 times faster, building a production-grade agent in days, not months. One favorite example is Asana, which built AI teammates so humans can collaborate directly with agents within Asana projects, delegating tasks. They built on Claude Managed Agents for speed and scale. Recently we upgraded it with multi-agent orchestration (build fleets of agents), outcomes (specify what success looks like and Claude iterates to get it done), and dreaming (Claude introspects on its previous transcripts to learn and self-improve). Today we're introducing two more: self-hosted sandboxes, so Claude can execute work on your own server, and MCP tunnels, so people who build internal agents can securely access their internal MCP servers. Claude Managed Agents spawn sandboxes when it's time to execute work like editing files; you can use ours or, starting today, your own — with first-class support for Daytona, Cloudflare, Vercel, and Modal sandboxes, with easy integration. With MCP tunnels, your MCP servers can be hosted behind your firewall on your private network so Claude Managed Agents can access them without exposing them over the public internet.
Caitlyn and I have been working with a fictional company called Counter, which offers software for small businesses to build a digital storefront. Counter wants to become an AI-native business, and one of its most common workflows is constantly running growth experiments to optimize merchant onboarding. So here we are in Counter's Slack workspace, with my colleague Gabe, wearing our cool Counter hats. We're excited to see how our latest growth experiment on the onboarding flow performed. Our new Growthbot came into the Slack channel and told us there's a clear winner: the simpler version of the onboarding flow is outperforming the older, longer one where merchants dropped off before finishing signup. The agent proactively tells us it's calling the experiment and getting started on cleaning up the old variant.
Let's look at how the agent is set up, using the Claude API CLI and the Claude developer console. First the agent configuration — system prompt, skills, tools. Then the MCP servers: Slack (the bot talked to us in Slack), a data warehouse so the bot can read experiment data, and a feature flags MCP server so the bot can call experiments or start new ones. Look at the URLs: Slack is a regular public MCP URL, but the data warehouse and feature-flags services are behind tunnel.anthropic.com — both MCP servers are behind Counter's firewall, only accessible via the tunnel, so Counter keeps its MCP servers on its private network. In the console, Counter set up a gateway in their own private network and established a secure connection to Anthropic, so any agent Caitlyn builds that uses the MCP tunnel can securely access internal MCP servers. You can create these tunnels directly in the console with a couple of configuration options. Counter's next requirement is that when the agent writes and executes code, it does so on a self-hosted sandbox, not public infrastructure. So for our managed agents we set up the environment to use a self-hosted sandbox — when the agent needs to execute work, it puts a work item in a queue, and here Counter uses Vercel's self-hosted sandbox, which picks up the item and spins up a sandbox in Counter's Vercel account to execute the work.
Managed agents have rich observability in the console. Let's check the session that produced that Slack thread. The agent runs a bunch of tools; we can confirm the things we cared about. There's a "set flag winner" tool call, meaning the agent worked through the MCP tunnel to call the feature-flag service. And when the agent writes the code to clean up the old variant, it checks out a new git branch and does this work on our Vercel self-hosted sandbox. Back in Slack, Growthbot is very proactive: it figured out it should default the simplified variant and put together a PR to do the defaulting and clean it up, executed on Counter's servers via the self-hosted sandbox. It even gave us a screenshot of the messy old variant versus the cleaner new one. It's also collaborative — when Caitlyn asked "What next?", Growthbot detected another opportunity to solve a 46% drop-off it found in the onboarding flow, calculated because it queried the data warehouse securely through an MCP tunnel, and asked if we wanted it to get started. Thank you, Claude. To wrap up: today we introduced two big upgrades to Claude Managed Agents — self-hosted sandboxes and MCP tunnels — making it easier than ever to become an AI-native business and build agents for all your workflows. Now Cat and Boris will talk about how Claude Code is making it even more fun to ship as a developer.
Angela and Caitlyn just showed how the Claude platform bridges the gap between what models can do and what businesses ship. We see a related challenge on Claude Code: bridging the gap between model capabilities and how every developer can benefit from them. First, thank you to all the builders in the room and watching online — thank you for trusting us and your production codebase back when Sonnet 3.7 was our frontier model and our product was rough around the edges. Our mission is to help every builder close the gap between having a great idea and having a shipped product. We do that by building tools that elicit the frontier intelligence from our models and making them accessible to every builder. We don't think of ourselves as having a finished roadmap — more like mountaineers climbing alongside you in terrain none of us has mapped, figuring things out as we go.
Last year I read every single edit Claude Code tried to make and every permission prompt, giving super-detailed guidance. Now most of us run in auto mode to delegate permissions, and we check in after Claude has tested, verified its changes, and has a PR ready. Claude Code started in the CLI, and the terminal is still the interface for power users who want minimal text and the most control. Then we added the IDE — the same agent, but you can follow along with code changes. Many of you now juggle multiple Claude Code instances, which we affectionately call multi-clauding. We added two new interfaces to manage more agents. One I use frequently is Claude Code on desktop — a full-screen graphical interface with built-in previews, a sidebar control plane, and the ability to render images and rich outputs. Desktop is a single view across local and cloud sessions, with visual indicators of which agents are running, blocked, or need input. Next is our newest surface, the Claude agents view in the CLI for people who prefer to stay in the terminal — see what's running, waiting, and done at a glance, reply inline to unblock, and jump in and out of sessions without losing your place. The VS Code extension and desktop app are both built on the Claude Agent SDK, the same one many of you build on. Many enterprises have adopted Claude Code wall-to-wall — at Anthropic it's driven a 200% increase in PRs per engineer even as our org has scaled substantially.
Here's some feedback we heard and what we built. You wanted to spend less time on code review, so we shipped a code review product that deploys a team of agents to traverse all your code changes and catch critical bugs — thousands of companies use it daily, including every internal Anthropic team. You wanted to code on the go, so we launched remote control in Claude Code on iOS and Android so you can fire off a task anywhere — go to a park, touch grass, and still get tasks done. You spend a lot of time babysitting PRs, fixing flaky CI tests, addressing review comments, and resolving merge conflicts, so we added autofix that listens for these events and proactively fixes each so your PR is always green. You wanted to run Claude Code on new tickets and customer bug reports, so we built routines: configure once and Claude Code runs on a schedule or in response to a webhook or API request. And you're landing so much code that security teams struggle to keep up, so we built Claude Security, which scans your codebase overnight, flags vulnerabilities ranked by severity, and lets you kick off Claude Code to address each. These primitives compose together. Companies have adopted Claude Code at the scale of entire engineering orgs: Shopify uses it across engineering and non-engineers (PMs, designers, data scientists) and brings it into their platform — Andrew McNamara, director of applied AI, says the speed is just crazy. Mercado Libre, Latin America's most popular e-commerce site with over 100 million buyers and 23,000 engineers, runs on Claude Code — it's reviewed more than 500,000 PRs with human oversight and modernized more than 9,000 apps, and Oscar (who leads technology) aims for 90% autonomous coding in a fully agent-driven PR loop by Q3. The detail I love most is that managers and VPs who haven't committed code in years are shipping again. Now, please welcome the creator of Claude Code, Boris Cherny.
I just realized we're matching. Thanks, Cat. Everything we're showing still feels magical to me — using Claude Code I get the same feelings I had programming my first calculator. Inside Anthropic, all day we're trading screenshots of things engineers are doing in the wild with Claude Code. For this demo, imagine we're an engineer at Acme Pay, a payment infrastructure company. We start in the Claude Code desktop app on one task: Claude is adding refunds to Acme's merchant dashboard — full implementation, idempotency so a duplicate webhook doesn't double-refund, multi-currency handling across regions, and audit logging for compliance. It writes the implementation and verifies its own work: it pulls up the merchant dashboard in the browser, triggers a refund, and the modal closes before the success toast appears — a real edge case. Claude sees the failure, traces it to a race condition in the optimistic update, fixes it, and verifies it works in the browser before calling the task done.
Now zoom out. Claude Code isn't running just one session — it's one of many running and managed in parallel. In the desktop app you can see all your Claude Code sessions: which are running, which need input, which have PRs merged and closed. Synchronous coding is now just a slice of what's happening, and going forward more code will be written async. That's why we keep talking about verification — if Claude can check its work, you can let it run while you work on something else and come back to a fully working result. A lot of my code these days is written by routines — I'm not doing the prompting, I create a routine that does the prompting. For engineers, think of it like a higher-order function: routines are a higher-order prompt. Back to the refund session — a teammate found a GitHub issue overnight, a routine watching the repo picked it up async and kicked off the work. With routines, developers set up async automations and wake up to PRs ready to merge. Routines can run on a schedule, be kicked off by webhooks or arbitrary API calls, and run locally or on remote cloud compute.
One more desktop feature: CI autofix. It watches the PR the last session opened and babysits it to production — auto-fixing comments from code review and security review, auto-fixing CI, and auto-rebasing on merge conflicts. Look what just happened: CI flicked on a network timeout, the routine woke up, diagnosed it as a known infra issue, and retried the job — and in the Claude Code codebase we tell it to fix the root cause instead of just retrying. CI is green, the engineer who owns the PR never sees the red X, and that work is off their plate. That's the shift: the default isn't "I'm going to prompt Claude Code," it's "I'm going to have Claude prompt Claude Code." Everything you just saw is available today, including routines and all the latest desktop updates. We hope these features let you close the gap between your ideas and shipping products.
That's what every talk today was pointing at — Lisa's capability curve, Angela and Caitlyn's agents that run on infrastructure you control, and what Cat and I just showed. These are three layers of one story: the capability is already here, and the remaining gap is how fast we put it to work. Spend the rest of today exploring these layers — research talks if you're evaluating models, Claude platform sessions if you're building for your users, and Claude Code workshops if you want more ways to bring Claude into your day-to-day workflow. Dive in, go deep, and start building with us. Thank you.