- Asana built "AI teammates" — agents that act as real actors in the system with sharing/RBAC controls, enterprise context, and shared memory — on top of Claude managed agents, generally available since March.
- The vision is "full multiplayer mode": multiple humans and agents collaborating on complex multi-step work (briefs, mock-ups, approvals) with auditability, rather than single-player agent use that doesn't compound knowledge.
- Claude managed agents handle the multi-step execution with built-in verification loops and a grader, letting Asana focus on its differentiators: the human interface layer, the work graph context, and security/guardrails.
Building with Claude Managed Agents and Asana AI teammates
- Asana's AI teammates are agents treated like onboarded human teammates: they have skills, full enterprise context, sharing and role-based access controls, and can work hand-in-hand with multiple people to complete multi-step work like approvals and end-to-end workflows.
- Most enterprises today use agents in a "single-player" way (one person gets an output, passes it on), which fails to build compounding knowledge or shared enterprise memory; Asana's goal is true multiplayer with multiple human-in-the-loop interactions.
- Enterprise memory is a key differentiator: agents accumulate historical interactions and context (e.g. a competitive-intelligence teammate built by ex-Asana employees that keeps improving as more people use it).
- Asana's 17-year "work graph" — mission/vision tracked by goals, in portfolios, delivered by projects, made of tasks with approvals and workflows — gives both humans (a UI) and agents (a way to represent themselves and get context) a shared structure, with security and guardrails preserved so agents don't leak information across projects.
- Claude managed agents are used specifically to complete multi-step actions: they reduced Asana's prototyping cost, come with a built-in verification loop, and include a grader that iterates over outcomes to ensure high-quality output.
- Compared to their previous messages-API approach, managed agents give faster prototyping (no manual agent loop, file management, or code execution to build), a much better verification process, and the ability to run multiple agents in parallel.
- Asana ships over 21 pre-built AI teammates matched to ideal customer profiles (PMO, marketing, IT, HR, R&D) that can do launch planning, spec writing, goal/resource management, pull data from Google Drive and Office 365, and produce slides, comments, and full HTML.
- In a demo, a marketer creates a Kanban task assigned to a teammate, which auto-loads relevant project context to write a campaign brief and generate an HTML landing-page mock-up; feedback given via comments (e.g. "make the primary color blue") is ingrained into the agent's memory so future marketers don't repeat the mistake, and the whole interaction is auditable and convertible to an approval.
Claude managed agents — Anthropic's managed agent capability that handles the agent loop, file management, code execution, verification, and grading so builders focus on their own logic.
AI teammate — Asana's term for an agent that is a real actor in the system with skills, context, sharing/RBAC controls, and the ability to collaborate with multiple humans.
Multiplayer mode — Multiple humans and agents collaborating on the same shared work, as opposed to single-player one-to-one agent use.
Enterprise memory — Persistent, shared context and historical interactions an agent accumulates, reusable across the team and compounding over time.
Work graph — Asana's 17-year-old structure linking mission/vision, goals, portfolios, projects, and tasks for both humans and agents.
Verification loop — A built-in process in managed agents that checks output quality before returning a result.
Grader — A component of managed agents that iterates over an outcome multiple times against a rubric to ensure a high-quality output.
RBAC / app actors — Role-based access controls treating agents as true actors in the system with defined sponsors and managers.
Outcome — The target result Asana passes to managed agents, including injected Asana context, which the grader evaluates against.
Hi everyone. I'm Arnauld from Asana and I'm here to talk to you about how we've built Asana's AI teammates on Claude managed agents. Our vision at Asana is to bring forward this promise of the agentic enterprise where human beings and AI agents can work together to get complex multi-step work done. Whether you're in IT or operations or product or marketing, our vision is the AI agent is an actor in the system. It's got a deep set of skills, it's got all of the context required, and it can work hand-in-hand with other human beings to get multi-step work done — things like approvals, complete end-to-end workflows, getting to real-world outcomes. AI teammates within Asana are generally available as of March.
What we see today in the enterprise is almost everybody is experimenting with AI agents. There's a lot of utilization of them. But most companies are still using AI agents in the single-player way, where an individual can interact with the agent, get an outcome, and then pass it on to someone else to complete those multi-step processes. You don't end up with compounding knowledge. You don't end up with this concept of a shared enterprise memory. You don't end up with concepts like true multiplayer, multiple human-in-the-loop interactions.
Our vision is full multiplayer mode. You have agents that are real actors in the system. They've got sharing and RBAC controls just like you would be onboarding a new human teammate into the system. They have the context, they can work with multiple people, they can get nudges and interactions with multiple people, and they can complete these end-to-end jobs to be done. One funny anecdote: there are a bunch of former Asanas who are now at Anthropic, and one of them had built an AI teammate for us that's a competitive intelligence researcher. Even though they're no longer at Asana, we still use it on a day-to-day basis, and all of the historical interactions that he had to set up the context and the knowledge — when we were doing competitive responses to RFPs — is ingrained in that agent, and it's just getting better and better as more people join the team and keep using it. That concept of enterprise memory and the shared ability for multiple people to use agents has been phenomenal for us internally.
The other concept we've been working on is this ability to ensure that the agent gets complete context about how your enterprise works: who does what by when and why, historical decisions, approvals, ways in which people have had back and forth on a particular campaign brief or project plan that have finally led to its approval. Those decisions are tracked inside Asana and we provide them to the teammate in a way that preserves security, guardrails, and auditability, so you can get real-world results and action.
Asana is the system of action for work. We have a work graph that we've been building for over 17 years. Every company has a mission and vision that is tracked by goals. Goals are in portfolios. Portfolios are delivered by a multitude of projects. Each project has a set of tasks, and those tasks could have approvals, workflows, and other things embedded in them. As you build out this context graph that is designed for both human beings and AI agents, human beings have an easy-to-comprehend UI to interact with how work gets done, and agents have a way to represent themselves and get the context they need to perform in a true multiplayer way, following the right security and guardrails required by your enterprise.
The place where we leverage Claude's managed agents capability is to complete those multi-step actions where Anthropic's tooling allows us to go complete the task. In the demo I'll showcase, you'll see a multi-step interaction between a marketer who's releasing a new campaign and multiple people on their team. That campaign requires them to not only write a brief, but also start producing a few mock-ups for what the landing page is going to look like so the team can get to a shared understanding of whether that's good to go. It involves creating a complex document using the right enterprise context, as well as creating multiple iterations of what the public website should look like — generating HTML. It's a multi-step action.
With the managed agents capability, there are a few things that make AI teammates far more powerful. It helped us reduce our prototyping costs — much faster prototyping for these actions and skills. It comes built-in with a verification loop, so we know the quality of the content at the end of the process is high. And it has a built-in grader: once Asana passes in the outcome it wants to Claude managed agents, the grader will iterate through that outcome multiple times to ensure we're getting a high-quality output. This allows us to focus on the things unique to Asana — the human interface layer to coordinate across multiple people, the system of context, the security — so that if you're using the same agent as me or Nigel or Tony and we're in different work graph objects or projects, the agent won't leak information across the other.
If I compare the way AI teammates work today with Claude managed agents versus the messages API we previously used: we're getting faster prototyping because we're not having to build a manual agent loop, file management, or code execution. The verification process has gotten way better with this built-in product. And we can enable multiple agents in parallel to work independently, because a lot of these knowledge-worker actions require multiple agents to work in parallel to produce a full plan and iterate through it.
As of today, there are over 21 pre-built agents called AI teammates within Asana. We designed the pre-built ones to match our ideal customer profiles across the office of the PMO, marketing, IT, HR, and R&D. They can help you with launch planning, writing specs, goal management, resource management and capacity planning. They use existing constructs within Asana for tracking portfolio goals and timeline updates, and they take real action — they can pull in data from Google Drive and Office 365 (with more integrations coming), produce slides, add comments, and create full HTML. A lot of those capabilities are possible because we've invested in managed agents, which allow us to run that verification loop and multi-step workflow.
Another example: we've been dogfooding this internally at Asana for months. I have an agent for the product management team — a "product thought buddy" — that has all the context around why we made certain trade-off decisions and what the strategy is. When our marketing team has questions like "we're planning Work Innovation Summit London, here's a draft of the keynote speech, could we get the product team's input?", they create a task in Asana and assign it to the product thought buddy. It has all the trade-offs, the current roadmap, and the context for how our roadmap works, and it produces a plan and feedback on that keynote speech that's highly optimized for the way we work and accurate from the perspective of using all the real-time context. It's done in a multiplayer way, so everybody in the product team can see that response, react to it, give it nudges, and it'll remember that across multiple runs.
Let's roll the demo video. The human being is a marketer trying to launch a new campaign, using AI teammates leveraging Claude's managed agents under the covers. They kick it off in a Kanban board inside Asana by creating a new task: "plan for a new campaign brief and prototype a landing page." Perhaps they haven't used AI teammates before, so they go to the teammate gallery and choose one of the pre-built teammates with all of their behavior guidance. When they customize it for their use, it automatically picks up the right work graph objects — previous campaign projects or portfolios — to add to that teammate's memory so it can produce better content. The teammate first plans out its response and creates a document, which is its campaign brief. It's also created an HTML file, which is the landing-page mock-up.
This shows what you'd see in the Claude console. We pass in the outcome to the managed agents product, and we can see the runs within the console to highlight that managed agents is actually running the grader to ensure the verification loop is complete and the outcome is appropriate. The combination has allowed our development team to ensure we're getting high-quality outcomes — delivering on the customer promise: you're getting all the security benefits of our RBAC controls and app actors (agents are true actors in the system), the enterprise context, and managed agents delivering on the outcome. They've gone from just an idea or task down to a page immediately.
Interacting with the agent and asking to iterate is as simple as talking to it via comments. Let's say it picked up green as the primary color, but the primary color has now changed. You can say "this is great, but let's make the primary color blue" and give it feedback that this is our new company primary color. This gets ingrained into the agent's memory, so if a different marketer picks up and uses that agent, it will not make that same mistake again — it will remember the new color scheme is blue. You can see the campaign brief writer is working on behalf of the admin, and the memory is getting created.
Now multiplayer in action: maybe the marketer feels this is a good starting point and wants to bring in someone else to review it. This new person comes in and says, "thank you for doing this, but I'd love to see an iteration that's more minimalistic in its look and feel." Just type it out as if it's a comment — and it is a real comment. Hit go, and the Claude managed agent functionality runs in the background and creates that minimalistic view. Because this is a shared workspace, everything that's happened — the interactions between humans and the AI agent — is tracked in the task and is fully auditable. At some point, if you need to bubble this up to your manager and get their approval, you can simply convert it to an approval. They'll see all the prompts that have gone in and the back and forth between the AI agent and the multiple humans, giving them confidence that the right questions have been answered.
In the teammate's detail page, there is also a concept of who owns or manages that AI agent. These RBAC controls mean there's a definition of the human beings who are the agent sponsors and who can manage it, so those humans get a chance to delete memories or change the parts of the Asana context graph the agent has access to, to ensure it continues to behave consistently going forward.
What you're seeing here is what we have generally available today: multi-step actions, documenting everything, working in a multiplayer way. The future for us is thinking about more capable, more multi-step workflows — entire launch-planning processes, resource management and capacity planning across hundreds of people, generating dynamic dashboards, risk reports, automatically alerting users on remediation steps. We're partnering with Anthropic on learning faster from team patterns and agent skills, and we're identifying opportunities to move work forward more proactively. Proactivity is something we're working on, where if an Asana AI teammate is part of a project or looking at your Kanban board, even if you haven't assigned it any task, it can automatically wake up and say "I can pick that up" or "I've seen this particular issue in a different project before, and this is how you resolve it."
This has allowed us as a company to focus on all the aspects that make Asana great: defining in a human-readable way what your multi-step processes are, and interacting with these agents in a way that is very human — it feels like onboarding a new person onto the team. You give them context with projects or documents, you give them feedback, and multiple people can use them. Even in my example of the person no longer at Asana, all of the work they've done and those memories are reusable by people on the team to keep doing that work faster and better going forward.
That was a primer on how we're using managed agents to deliver Asana AI teammates. Let me take some questions.
Q: You mentioned the validation of the agent happens on the Anthropic side. Is Asana domain knowledge being injected into the verification process? At a high level, we are passing in the Asana context as part of the outcome definition. As we test it internally before releasing a pre-built capability to customers, there's another level of QA within Asana by our engineering team. And because we're a very human-in-the-loop system, if there's any issue with the quality of the output for a particular customer — perhaps because their context isn't fully fleshed out — the more you use it and the more nudges you give it across multiple people, all of that compounds and keeps getting better, because we record that back within Asana and push it back as part of the call to run the managed agent every single time.
Q: How did you design the rubric that your grader sees to iterate on outputs? Any learnings? We want to approach it like we approach any other prompt. In the end, all the text you provide to the model is a different form of prompt, and you want to have different evaluations for the outcomes for different things you're trying to achieve. If you're able to instrument that just like anything else, you're able to iterate quickly and make confident decisions.
Q: How are you thinking about skill maintenance over time? We're a product for knowledge workers, so we want to make it as simple as possible for customers to deploy and use this. Asana decides what skills get baked into the generally available product, working backwards from our ideal customer profiles and the advances in the underlying model provider. That allows us to be very prescriptive about the additional capabilities available in a GA way. It's a shrink-wrapped product. Over time we might open it up so customers can design their own skills and highly customize their AI teammates, but the primary go-to-market motion is that we pre-design these AI teammates, so we control the quality level, which ones get released, and the lifecycle on the R&D side.
Q: How do you manage third-party integrations — at the Asana level or the agent level? In terms of third-party integrations and managed agents, we integrate them at both levels. We integrate them directly with our own AI teammates agent loop, and we also integrate them at the MCP level with managed agents. We want to make sure that the managed agents, when we're using them, have all of the context they need, and that we're also supplying context when we're not using managed agents.
Thanks everybody for joining us. If you have a chance, give AI teammates a try. Thank you.
TL;DR
- Asana built "AI teammates" — agents that act as real actors in the system with sharing/RBAC controls, enterprise context, and shared memory — on top of Claude managed agents, generally available since March.
- The vision is "full multiplayer mode": multiple humans and agents collaborating on complex multi-step work (briefs, mock-ups, approvals) with auditability, rather than single-player agent use that doesn't compound knowledge.
- Claude managed agents handle the multi-step execution with built-in verification loops and a grader, letting Asana focus on its differentiators: the human interface layer, the work graph context, and security/guardrails.
Takeaways
- Asana's AI teammates are agents treated like onboarded human teammates: they have skills, full enterprise context, sharing and role-based access controls, and can work hand-in-hand with multiple people to complete multi-step work like approvals and end-to-end workflows.
- Most enterprises today use agents in a "single-player" way (one person gets an output, passes it on), which fails to build compounding knowledge or shared enterprise memory; Asana's goal is true multiplayer with multiple human-in-the-loop interactions.
- Enterprise memory is a key differentiator: agents accumulate historical interactions and context (e.g. a competitive-intelligence teammate built by ex-Asana employees that keeps improving as more people use it).
- Asana's 17-year "work graph" — mission/vision tracked by goals, in portfolios, delivered by projects, made of tasks with approvals and workflows — gives both humans (a UI) and agents (a way to represent themselves and get context) a shared structure, with security and guardrails preserved so agents don't leak information across projects.
- Claude managed agents are used specifically to complete multi-step actions: they reduced Asana's prototyping cost, come with a built-in verification loop, and include a grader that iterates over outcomes to ensure high-quality output.
- Compared to their previous messages-API approach, managed agents give faster prototyping (no manual agent loop, file management, or code execution to build), a much better verification process, and the ability to run multiple agents in parallel.
- Asana ships over 21 pre-built AI teammates matched to ideal customer profiles (PMO, marketing, IT, HR, R&D) that can do launch planning, spec writing, goal/resource management, pull data from Google Drive and Office 365, and produce slides, comments, and full HTML.
- In a demo, a marketer creates a Kanban task assigned to a teammate, which auto-loads relevant project context to write a campaign brief and generate an HTML landing-page mock-up; feedback given via comments (e.g. "make the primary color blue") is ingrained into the agent's memory so future marketers don't repeat the mistake, and the whole interaction is auditable and convertible to an approval.
Vocabulary
Claude managed agents — Anthropic's managed agent capability that handles the agent loop, file management, code execution, verification, and grading so builders focus on their own logic.
AI teammate — Asana's term for an agent that is a real actor in the system with skills, context, sharing/RBAC controls, and the ability to collaborate with multiple humans.
Multiplayer mode — Multiple humans and agents collaborating on the same shared work, as opposed to single-player one-to-one agent use.
Enterprise memory — Persistent, shared context and historical interactions an agent accumulates, reusable across the team and compounding over time.
Work graph — Asana's 17-year-old structure linking mission/vision, goals, portfolios, projects, and tasks for both humans and agents.
Verification loop — A built-in process in managed agents that checks output quality before returning a result.
Grader — A component of managed agents that iterates over an outcome multiple times against a rubric to ensure a high-quality output.
RBAC / app actors — Role-based access controls treating agents as true actors in the system with defined sponsors and managers.
Outcome — The target result Asana passes to managed agents, including injected Asana context, which the grader evaluates against.
Transcript
Hi everyone. I'm Arnauld from Asana and I'm here to talk to you about how we've built Asana's AI teammates on Claude managed agents. Our vision at Asana is to bring forward this promise of the agentic enterprise where human beings and AI agents can work together to get complex multi-step work done. Whether you're in IT or operations or product or marketing, our vision is the AI agent is an actor in the system. It's got a deep set of skills, it's got all of the context required, and it can work hand-in-hand with other human beings to get multi-step work done — things like approvals, complete end-to-end workflows, getting to real-world outcomes. AI teammates within Asana are generally available as of March.
What we see today in the enterprise is almost everybody is experimenting with AI agents. There's a lot of utilization of them. But most companies are still using AI agents in the single-player way, where an individual can interact with the agent, get an outcome, and then pass it on to someone else to complete those multi-step processes. You don't end up with compounding knowledge. You don't end up with this concept of a shared enterprise memory. You don't end up with concepts like true multiplayer, multiple human-in-the-loop interactions.
Our vision is full multiplayer mode. You have agents that are real actors in the system. They've got sharing and RBAC controls just like you would be onboarding a new human teammate into the system. They have the context, they can work with multiple people, they can get nudges and interactions with multiple people, and they can complete these end-to-end jobs to be done. One funny anecdote: there are a bunch of former Asanas who are now at Anthropic, and one of them had built an AI teammate for us that's a competitive intelligence researcher. Even though they're no longer at Asana, we still use it on a day-to-day basis, and all of the historical interactions that he had to set up the context and the knowledge — when we were doing competitive responses to RFPs — is ingrained in that agent, and it's just getting better and better as more people join the team and keep using it. That concept of enterprise memory and the shared ability for multiple people to use agents has been phenomenal for us internally.
The other concept we've been working on is this ability to ensure that the agent gets complete context about how your enterprise works: who does what by when and why, historical decisions, approvals, ways in which people have had back and forth on a particular campaign brief or project plan that have finally led to its approval. Those decisions are tracked inside Asana and we provide them to the teammate in a way that preserves security, guardrails, and auditability, so you can get real-world results and action.
Asana is the system of action for work. We have a work graph that we've been building for over 17 years. Every company has a mission and vision that is tracked by goals. Goals are in portfolios. Portfolios are delivered by a multitude of projects. Each project has a set of tasks, and those tasks could have approvals, workflows, and other things embedded in them. As you build out this context graph that is designed for both human beings and AI agents, human beings have an easy-to-comprehend UI to interact with how work gets done, and agents have a way to represent themselves and get the context they need to perform in a true multiplayer way, following the right security and guardrails required by your enterprise.
The place where we leverage Claude's managed agents capability is to complete those multi-step actions where Anthropic's tooling allows us to go complete the task. In the demo I'll showcase, you'll see a multi-step interaction between a marketer who's releasing a new campaign and multiple people on their team. That campaign requires them to not only write a brief, but also start producing a few mock-ups for what the landing page is going to look like so the team can get to a shared understanding of whether that's good to go. It involves creating a complex document using the right enterprise context, as well as creating multiple iterations of what the public website should look like — generating HTML. It's a multi-step action.
With the managed agents capability, there are a few things that make AI teammates far more powerful. It helped us reduce our prototyping costs — much faster prototyping for these actions and skills. It comes built-in with a verification loop, so we know the quality of the content at the end of the process is high. And it has a built-in grader: once Asana passes in the outcome it wants to Claude managed agents, the grader will iterate through that outcome multiple times to ensure we're getting a high-quality output. This allows us to focus on the things unique to Asana — the human interface layer to coordinate across multiple people, the system of context, the security — so that if you're using the same agent as me or Nigel or Tony and we're in different work graph objects or projects, the agent won't leak information across the other.
If I compare the way AI teammates work today with Claude managed agents versus the messages API we previously used: we're getting faster prototyping because we're not having to build a manual agent loop, file management, or code execution. The verification process has gotten way better with this built-in product. And we can enable multiple agents in parallel to work independently, because a lot of these knowledge-worker actions require multiple agents to work in parallel to produce a full plan and iterate through it.
As of today, there are over 21 pre-built agents called AI teammates within Asana. We designed the pre-built ones to match our ideal customer profiles across the office of the PMO, marketing, IT, HR, and R&D. They can help you with launch planning, writing specs, goal management, resource management and capacity planning. They use existing constructs within Asana for tracking portfolio goals and timeline updates, and they take real action — they can pull in data from Google Drive and Office 365 (with more integrations coming), produce slides, add comments, and create full HTML. A lot of those capabilities are possible because we've invested in managed agents, which allow us to run that verification loop and multi-step workflow.
Another example: we've been dogfooding this internally at Asana for months. I have an agent for the product management team — a "product thought buddy" — that has all the context around why we made certain trade-off decisions and what the strategy is. When our marketing team has questions like "we're planning Work Innovation Summit London, here's a draft of the keynote speech, could we get the product team's input?", they create a task in Asana and assign it to the product thought buddy. It has all the trade-offs, the current roadmap, and the context for how our roadmap works, and it produces a plan and feedback on that keynote speech that's highly optimized for the way we work and accurate from the perspective of using all the real-time context. It's done in a multiplayer way, so everybody in the product team can see that response, react to it, give it nudges, and it'll remember that across multiple runs.
Let's roll the demo video. The human being is a marketer trying to launch a new campaign, using AI teammates leveraging Claude's managed agents under the covers. They kick it off in a Kanban board inside Asana by creating a new task: "plan for a new campaign brief and prototype a landing page." Perhaps they haven't used AI teammates before, so they go to the teammate gallery and choose one of the pre-built teammates with all of their behavior guidance. When they customize it for their use, it automatically picks up the right work graph objects — previous campaign projects or portfolios — to add to that teammate's memory so it can produce better content. The teammate first plans out its response and creates a document, which is its campaign brief. It's also created an HTML file, which is the landing-page mock-up.
This shows what you'd see in the Claude console. We pass in the outcome to the managed agents product, and we can see the runs within the console to highlight that managed agents is actually running the grader to ensure the verification loop is complete and the outcome is appropriate. The combination has allowed our development team to ensure we're getting high-quality outcomes — delivering on the customer promise: you're getting all the security benefits of our RBAC controls and app actors (agents are true actors in the system), the enterprise context, and managed agents delivering on the outcome. They've gone from just an idea or task down to a page immediately.
Interacting with the agent and asking to iterate is as simple as talking to it via comments. Let's say it picked up green as the primary color, but the primary color has now changed. You can say "this is great, but let's make the primary color blue" and give it feedback that this is our new company primary color. This gets ingrained into the agent's memory, so if a different marketer picks up and uses that agent, it will not make that same mistake again — it will remember the new color scheme is blue. You can see the campaign brief writer is working on behalf of the admin, and the memory is getting created.
Now multiplayer in action: maybe the marketer feels this is a good starting point and wants to bring in someone else to review it. This new person comes in and says, "thank you for doing this, but I'd love to see an iteration that's more minimalistic in its look and feel." Just type it out as if it's a comment — and it is a real comment. Hit go, and the Claude managed agent functionality runs in the background and creates that minimalistic view. Because this is a shared workspace, everything that's happened — the interactions between humans and the AI agent — is tracked in the task and is fully auditable. At some point, if you need to bubble this up to your manager and get their approval, you can simply convert it to an approval. They'll see all the prompts that have gone in and the back and forth between the AI agent and the multiple humans, giving them confidence that the right questions have been answered.
In the teammate's detail page, there is also a concept of who owns or manages that AI agent. These RBAC controls mean there's a definition of the human beings who are the agent sponsors and who can manage it, so those humans get a chance to delete memories or change the parts of the Asana context graph the agent has access to, to ensure it continues to behave consistently going forward.
What you're seeing here is what we have generally available today: multi-step actions, documenting everything, working in a multiplayer way. The future for us is thinking about more capable, more multi-step workflows — entire launch-planning processes, resource management and capacity planning across hundreds of people, generating dynamic dashboards, risk reports, automatically alerting users on remediation steps. We're partnering with Anthropic on learning faster from team patterns and agent skills, and we're identifying opportunities to move work forward more proactively. Proactivity is something we're working on, where if an Asana AI teammate is part of a project or looking at your Kanban board, even if you haven't assigned it any task, it can automatically wake up and say "I can pick that up" or "I've seen this particular issue in a different project before, and this is how you resolve it."
This has allowed us as a company to focus on all the aspects that make Asana great: defining in a human-readable way what your multi-step processes are, and interacting with these agents in a way that is very human — it feels like onboarding a new person onto the team. You give them context with projects or documents, you give them feedback, and multiple people can use them. Even in my example of the person no longer at Asana, all of the work they've done and those memories are reusable by people on the team to keep doing that work faster and better going forward.
That was a primer on how we're using managed agents to deliver Asana AI teammates. Let me take some questions.
Q: You mentioned the validation of the agent happens on the Anthropic side. Is Asana domain knowledge being injected into the verification process? At a high level, we are passing in the Asana context as part of the outcome definition. As we test it internally before releasing a pre-built capability to customers, there's another level of QA within Asana by our engineering team. And because we're a very human-in-the-loop system, if there's any issue with the quality of the output for a particular customer — perhaps because their context isn't fully fleshed out — the more you use it and the more nudges you give it across multiple people, all of that compounds and keeps getting better, because we record that back within Asana and push it back as part of the call to run the managed agent every single time.
Q: How did you design the rubric that your grader sees to iterate on outputs? Any learnings? We want to approach it like we approach any other prompt. In the end, all the text you provide to the model is a different form of prompt, and you want to have different evaluations for the outcomes for different things you're trying to achieve. If you're able to instrument that just like anything else, you're able to iterate quickly and make confident decisions.
Q: How are you thinking about skill maintenance over time? We're a product for knowledge workers, so we want to make it as simple as possible for customers to deploy and use this. Asana decides what skills get baked into the generally available product, working backwards from our ideal customer profiles and the advances in the underlying model provider. That allows us to be very prescriptive about the additional capabilities available in a GA way. It's a shrink-wrapped product. Over time we might open it up so customers can design their own skills and highly customize their AI teammates, but the primary go-to-market motion is that we pre-design these AI teammates, so we control the quality level, which ones get released, and the lifecycle on the R&D side.
Q: How do you manage third-party integrations — at the Asana level or the agent level? In terms of third-party integrations and managed agents, we integrate them at both levels. We integrate them directly with our own AI teammates agent loop, and we also integrate them at the MCP level with managed agents. We want to make sure that the managed agents, when we're using them, have all of the context they need, and that we're also supplying context when we're not using managed agents.
Thanks everybody for joining us. If you have a chance, give AI teammates a try. Thank you.