Enterprise groups constructing AI brokers preserve hitting the identical wall: a chatbot that may reply a immediate however can't keep in mind what the final 5 individuals requested it, and might't inform you whether or not final month's model truly labored.
In a hearth chat with VentureBeat's Sam Witteveen at VB Transform 2026, Asana's chief product officer, Arnab Bose, unpacked how his crew tackled this downside to construct a brand new working system: Agentic Work Administration (AWM). The product treats AI brokers as coachable teammates that function alongside people slightly than as one-to-one assistants.
For product builders and builders making an attempt to maneuver past primary integrations, Bose offered a glance beneath the hood. He detailed how Asana engineered AWM, providing a blueprint for fixing real-world bottlenecks and constructing agentic methods at scale.
The Work Graph: 18 years of firm knowledge, repurposed
To construct an working system for human-agent groups, Asana wanted a ready-made enterprise context graph. They constructed AWM on prime of their 18-year-old structure: the Work Graph.
This graph-based database organizes info via a construction the corporate calls the Pyramid of Readability. The smallest unit of labor is a activity with an assignee and a due date. Duties belong to initiatives, initiatives roll up into portfolios, and portfolios hook up with company-wide targets. The graph might help hint for instance how a delayed design activity impacts a company income objective. The Work Graph supplies a real-time ledger of who does what, by when, and why.
AWM leverages this structure to create a multiplayer teammate. A regular AI copilot is stateless and tied to a single consumer's immediate. As a result of AWM plugs into the Work Graph, the AI can view overarching firm targets, replace challenge statuses, and share reminiscence with human colleagues.
"As a result of [the agent] is plugged into the Work Graph, it's not simply a specific immediate that you just're sending it or a specific particular person's markdown file system on their native file,” Bose mentioned. “It's working off of that shared ledger for the entire firm."
AWM is already in manufacturing. Bose mentioned Asana has "a number of clients reside and profitable on it," together with FedEx, which printed its personal case examine on the shift.
Constructing in guardrails for confidential work
Delivery AWM to enterprise clients required Asana to resolve a number of technical hurdles. The primary was knowledge governance. If an AI teammate acts throughout an organization, it builds a shared reminiscence by studying from workflows and human suggestions.
Bose highlighted a crucial boundary downside: If an govt makes use of AWM to construct workflows for a confidential challenge, the system should make sure the agent's up to date reminiscence doesn’t leak context to an unauthorized worker who interacts with the identical agent later.
"[I] shouldn't be capable to leverage that shared reminiscence once I run the AI teammate should you created that reminiscence utilizing that very same teammate on a challenge that’s, let's say, a secret M&A challenge that I don't have entry to," Bose mentioned. Asana engineered a system of entry controls to control what triggers the creation of a reminiscence versus the straightforward execution of a activity.
Second, AWM handles dynamic mannequin routing to summary immediate engineering away from the consumer. When a consumer assigns a activity to an AI teammate (i.e., drafting a job description for a common supervisor function), the AI cross-references public job postings, Asana’s inside type information, and product requirement paperwork. For a fancy activity, the system mechanically routes the immediate to a heavy frontier mannequin — Bose pointed to Anthropic's Opus and OpenAI's fashions as examples — whereas lighter duties get down-leveled to one thing quicker and cheaper.
"We don't need the data employee to should suppose via what the absolute best immediate, context engineering, and attachments are that they need to put into the duty," Bose mentioned. "It ought to really feel as should you had been assigning the duty to a human being."
This dynamic routing introduces a 3rd problem: billing abstraction. Agentic duties range in computational complexity, making credit score burn charges unpredictable.
"We don't wish to get right into a state the place our clients are having to cause about the truth that a few of these duties… are far more advanced than others they usually'll be burning credit at completely different charges," Bose mentioned, including that unpredictable pricing risked clients throttling their very own workers by capping how typically they might run an AI teammate.
To make AWM commercially viable, Asana designed its billing structure to cost a static price per activity completion. The platform absorbs the complexity of mannequin choice, token counts, and run limits to make sure predictable enterprise pricing.
The issue with stateless chatbots
AWM targets a selected downside with present enterprise AI deployments: statelessness. Builders can simply join giant language fashions to enterprise instruments like Slack, Google Drive, or Databricks utilizing Mannequin Context Protocol (MCP) integrations. Nevertheless, primary chat-based brokers lack persistence.
Bose detailed a state of affairs the place a consumer asks a chat agent to draft a advertising and marketing marketing campaign primarily based on historic efficiency and aggressive analysis. The agent fetches knowledge from exterior instruments to reply the immediate, however the execution occurs in a vacuum. It’s a one-off activity that advantages a single particular person. It fails to create a reusable workflow for the following individual constructing an analogous marketing campaign.
"The problem with that’s that these calls are stateless, and they don’t seem to be leveraging a shared firm mind that’s this graph-based database or a context graph," Bose mentioned.
AWM solves this by making a everlasting state. When an AI teammate inside AWM completes a activity, the system data the metadata. It registers whether or not the completion improved the challenge standing and the way it moved higher-level firm targets.
Inside CoreWeave's product launches
Cloud supplier CoreWeave is an early adopter utilizing AWM to overtake advanced new product launches.
"CoreWeave is utilizing each our deterministic AI studio workflow guidelines in addition to a number of AI teammates to do new product launches," Bose shared.
Prior to now, CoreWeave product managers stuffed out difficult kinds detailing infrastructure, parameters, and prices. Human reviewers manually evaluated these kinds and broke them out into particular duties for finance, advertising and marketing, and {hardware} groups.
Underneath the AWM workflow, a product supervisor writes a typical Google doc pointing to their product requirement paperwork. A deterministic AI workflow reads the doc, mechanically creates the challenge construction, and assigns duties. Specialised brokers then take over the execution. One agent then watches total challenge standing and flags bottlenecks; one other, working inside particular person duties, forecasts infrastructure prices and recommends approvals when the numbers align with historic budgets. The system mechanically triages the busywork whereas human beings concentrate on evaluating the AI's outputs.
The frenemy downside
The dynamic will get difficult by the truth that the identical frontier-model suppliers powering AWM beneath the hood — Anthropic, OpenAI — are additionally delivery their very own competing agent merchandise, like Anthropic's Claude in Slack (Tag). Pressed on the overlap, Bose didn't dispute the strain.
"I feel that's the fact that all of us should reside in," he mentioned.
His case for AWM's endurance rests on Asana's 18 years of user-experience and workflow knowledge, and prebuilt normal working procedures for particular industries — experience he argues uncooked frontier fashions don't have. A product like Tag can work effectively in Slack, he mentioned, nevertheless it requires a extremely curated channel and its personal separate credentials for each downstream app it touches.
"There's a giant distinction between the facility of the mannequin plus a light-weight option to reveal its worth, and one thing that's pre-built … for true end-to-end use," Bose mentioned.
