AI
The Single-Player AI Trap
Most enterprises use AI agents in single-player mode: one person gets an output, then passes it to the next. That pattern breaks down for complex, multi-step work. There’s no compounding knowledge, no shared enterprise memory, and no true multi-human interaction. Arnab Bose opened by naming what’s missing: agents that accumulate context across teams, remember prior decisions, and collaborate with multiple people simultaneously rather than serving as isolated one-shot tools.
Asana’s Work Graph as Enterprise Memory
Asana’s work graph maps a company’s operating structure: missions track to goals, goals to portfolios, portfolios to projects, projects to tasks with embedded approvals and workflows. That 17-year investment in codifying how organizations work is what agents read for context. When an AI teammate joins a project, it gets historical decisions, approval chains, and team communication. It also gets security guardrails. An agent in one project won’t leak data from another.
What Claude Managed Agents Replaced
Before managed agents, the Asana team hand-rolled agent loops, file management, and code execution. The switch to Claude managed agents cut prototyping time. More important: it added a built-in verification loop with a grader that iterates on outputs until they meet the defined outcome. Multiple agents can now run in parallel, which matters for knowledge-worker tasks that require several simultaneous work streams. Asana passes the outcome spec to managed agents, and the grader handles quality assurance.
21 Pre-Built Teammates Across Business Functions
As of the talk, Asana ships 21 pre-built AI teammates covering PMO, marketing, IT, HR, and R&D. Each is pre-configured around that department’s job-to-be-done: launch planning, spec writing, goal management, resource planning. They pull data from Google Drive and Office 365 and act in those systems: adding comments, producing slides, generating HTML. One internal example: a competitive intelligence agent built by a former employee still runs daily, getting sharper as more team members interact with it.
Multiplayer Mode and Compounding Memory
The demo shows a marketer requesting a campaign brief and landing page HTML from a Kanban task. After the first draft, a second team member asks for a minimalist iteration via a comment. The agent processes both requests, and the exchange logs as a fully auditable record. When the marketer corrects the primary color from green to blue, the preference gets stored in the agent’s memory. The next marketer who uses that agent won’t repeat the mistake. Managers reviewing the task see every prompt and response before approving.
Q&A
How does Asana domain knowledge get into the managed agents verification loop? Asana passes context as part of the outcome definition; engineering QA runs before GA release, and quality compounds as teams add nudges over multiple runs. ▶ 18:11
How should teams approach writing rubrics for the grader? Treat rubric text like any other prompt, define separate evaluations per outcome type, and instrument for fast iteration and confident decisions. ▶ 20:10
How does Asana manage skill maintenance over time? Skills are pre-designed and “shrink wrapped” around customer job profiles; Asana controls the release lifecycle, with potential for customer customization later. ▶ 21:43
How does the Asana team handle third-party integrations with managed agents? Integrations happen at both the AI teammates agent loop level and at the MCP level with managed agents directly. ▶ 23:41
Notable Quotes
You don’t end up with compounding knowledge. Arnab Bose · ▶ 1:30
you’ve got like a blue uh campaign brief landing page. Arnab Bose · ▶ 13:41
I I think we want to approach it like we approach any other prompt. Arnab Bose · ▶ 21:02
Key Takeaways
- Claude managed agents replaced hand-rolled loops, adding a built-in grader for verified multi-step outputs.
- Asana agents act as persistent, permissioned actors in the work graph, accumulating memory across all team interactions.
- Multiplayer mode lets multiple humans direct the same agent; every exchange is fully auditable before manager approval.