The key to successful AI agents within the enterprise? Shared memory and context.
This, according to Asana CPO Arnab Bose, provides detailed history and direct access from the destination — with guardrail checkpoints and human oversight, of course.
This way, “when you delegate, you don’t have to go ahead and redraw the whole context about how your business works,” Bose said at a recent VB event in San Francisco.
AI as an active partner, rather than a passive supplement
Asana launched Asana AI Teammates last year with the philosophy that, just like humans, AI agents should be connected to a team or project to create a collaborative system. To further this project, the project management company is fully integrated with Anthropic’s Claude.
Users can choose from 12 pre-built agents – for common use cases such as IT ticket deviation – or create their own, then assign them to project teams and quickly provide a historical record of which tasks have already been completed and what is yet to be resolved. Agents also have access to third-party services such as Microsoft 365 or Google Drive.
“When that agent is created, it doesn’t represent someone, it identifies itself as a partner and it gets all the same sharing permissions, you get that,” Bose explained. Everything anyone does — humans and AI included — is documented to allow for “easy clarification” and a “transparent and reliable system.”
But like human workers, AI agents are kept on standby: Importantly, the workflow includes checkpoints, where people can provide feedback and ask the agent to adjust certain aspects of the project or adjust research plans. This is written in what Bose calls “the most readable form of man.”
And importantly, the UI provides instructions and information about the agent’s behavior, and authorized administrators can pause, schedule and redirect models in the API when they perform actions based on conflicting directions or start acting “in a strange way.”
“Someone with editing rights can remove those conflicting elements and bring it back to its proper behavior,” said Bose. “We rely on that common pattern of interaction with people.”
Overcoming the challenges of accreditation, integration
But because AI agents are new, there are still many challenges in terms of security, accessibility and compatibility.
Asana users, for example, must go through the OAuth flow and grant Claude access to Asana through their MCP and other public APIs. But making all information workers aware that such integration exists – and more importantly, which OAuth grants are SAFE and should be avoided – can be a tall order.
Other challenges around direct OAuth grants between applications may be centralized by identity providers, noted Bose, or a centralized list of authorized AI business agents and their skill sets, “almost like an active directory or a directory of universal agents.”
Currently, however, beyond what Asana does, there is no standard process around shared information and memory, Bose said. His team has been getting “a lot of interesting internal inquiries” from partners who want their agents to work on Asana’s task graph and benefit from shared work.
“But because there is no protocol or standard, today it has to be a customized conversation,” said Bose.
Finally, there are three questions that CPO calls “extremely interesting” in AI music right now:
-
How do you build, manage and secure a whitelist of authorized AI agents?
-
How can you enable application-to-application integration as an IT team without stopping malicious or malicious agents?
-
Today’s agent-to-agent interactions are very one-player. Clouds can be independently connected to Asana or Figma or Slack. How can we ultimately achieve a unified, multiplayer outcome?
The increasing adoption of the modern protocol (MCP) – an open standard introduced by Anthropic that connects AI agents and external systems in a single action, instead of custom integration in every single pairing – is promising, he noted, and its widespread adoption could open up new and exciting use cases.
However, “I think there is probably no silver bullet level yet,” said Bose.
