The key to successful AI agents within an enterprise? Shared memory and context.
According to Asan CPO Arnab Bose, it offers detailed history and direct access from the beginning — with guardrail checkpoints and human oversight, of course.
That way, “when you delegate a task, you don’t have to go ahead and provide all the context again about how your business works,” Bose said at a recent VB event in San Francisco.
AI as an active teammate rather than a passive add-on
Asana launched Asana AI Teammates last year with the philosophy that like humans, AI agents should be paired directly into a team or project to create a collaborative system. To further this mission, the project management company has fully integrated with Anthropic’s cloud.
Users can choose from 12 pre-built agents – for common use cases like IT ticket deflection – or create their own, then assign them to project teams and instantly provide a historical record of which tasks have already been completed and what still needs to be resolved. Agents also have access to third-party resources like Microsoft 365 or Google Drive.
“When that agent is created, he’s not acting on behalf of anyone, he identifies himself as a teammate and he gets all the same sharing permissions, he inherits that,” Bose explained. Everything anyone does – including humans and AI – is documented to allow for “ease of explaining” and a “very transparent and trustworthy system”.
But like human workers, AI agents are kept under control: Critically, workflows include checkpoints where humans can respond and ask the agent to change certain elements of a project or adjust research plans. This has been documented by Bose in a “very human-readable way”.
Also important, the UI provides instructions and knowledge about the agent’s behavior, and approved administrators can pause, edit, and redirect models in the API when they take actions based on conflicting instructions or start acting “strangely.”
“A person with editing rights can remove things that are contradictory and bring it back to its correct behavior,” Bose said. “We’re leaning toward that normal human-comprehensible interaction pattern.”
Overcoming the challenges of authorization and integration
But since AI agents are so new, there are still many challenges around security, accessibility, and compatibility.
For example, Asana users must go through an OAuth flow and the cloud must grant access to Asana through its MCP and other public APIs. But telling all knowledge workers that that integration exists – and more importantly, which OAuth grants are OK and which should be avoided – can be a daunting task.
Some of the challenges around direct OAuth grants between applications could be addressed by centralization by identity providers, Bose said, or a centralized list of approved enterprise AI agents with their skill sets, “almost like an active directory or universal directory of agents.”
Bose said, however, that beyond what Asana is doing, there is no standard protocol around shared knowledge and memory. His team is getting “a lot of interesting inbound queries” from partners who want their agents to work on the Asana task graph and benefit from shared work.
“But since the protocol or standard doesn’t exist, it has to be a very complex custom interaction today,” Bose said.
Ultimately, there are three questions in AI orchestration right now that the CPO called “extremely interesting”:
- How do you create, manage, and secure an authoritative list of known approved AI agents?
-
How can you as an IT team enable app-to-app integration without configuring potentially dangerous or harmful agents?
-
Today’s agent-to-agent interactions are very much single-player. Clouds can be freely linked to Asana or Figma or Slack. How can we ultimately achieve a unified, multiplayer outcome?
The adoption of the Modern Reference Protocol (MCP) — an open standard introduced by Anthropic that connects AI agents to external systems in a single action rather than custom integration for each pair — is promising, he said, and its widespread adoption could open up new and exciting use cases.
However, “I think there’s probably no silver bullet standard out there right now,” Bose said.
<a href