Asana built AI agents that share memory across your whole company, with a wall around the secrets

Key points

• Asana launched Agentic Work Management (AWM), a system where AI agents work as shared teammates across an entire company rather than as personal assistants tied to one user.
• AWM is built on Asana's 18-year-old Work Graph database, which links individual tasks to company-wide goals in a single shared ledger.
• FedEx and CoreWeave are among the early enterprise customers already running AWM in production.
• Asana engineered access controls so an AI agent cannot pass knowledge from a confidential project to an employee who lacks permission to see that project.
• AWM charges a fixed price per task so companies can predict costs, even when some tasks require far more AI computing power than others.

Most AI assistants have a serious memory problem. A chatbot can draft a marketing plan on request. Come back the next morning and ask what it learned from yesterday's work, and it draws a blank. Every conversation starts from zero.

Asana's answer is a product it calls Agentic Work Management, or AWM. The project-management software company unveiled how it works at the VB Transform 2026 conference, in a talk first reported by VentureBeat. CoreWeave, one of the enterprise customers already running AWM, featured in our 23 July story on how most teams are measuring their AI agents the wrong way.

How does AWM actually work?

AWM treats AI agents, software that can carry out multi-step tasks on its own, as permanent members of a team rather than temporary helpers. The key is the Work Graph, a database Asana has been building for 18 years that connects every task to a project, every project to a portfolio, and every portfolio to a company goal.

When an AI agent inside AWM finishes a task, it doesn't just return an answer and disappear. The system records what happened, whether the work moved the project forward, and how it affected higher-level targets. That record becomes part of a shared memory that any human or AI teammate in the company can draw on later.

"It's working off of that shared ledger for the whole company," Asana chief product officer Arnab Bose said at the event.

What stops the AI from leaking confidential information?

Shared memory creates an obvious risk. An executive who uses the AI agent to plan a secret acquisition has taught that agent something sensitive, and a junior employee asking the same agent for help later shouldn't be able to benefit from it.

Asana engineered access controls that govern when the system creates a memory versus when it simply executes a task. The agent checks whether the person making a request has permission to see the project that originally generated the relevant knowledge. If they don't, that knowledge stays invisible.

Bose was direct: "[I] shouldn't be able to use that shared memory when I run the AI teammate if you created that memory using that same teammate on a project that is, let's say, a secret M&A project that I don't have access to."

AWM also handles model selection automatically. A large language model, the technology behind chatbots like ChatGPT and Claude, comes in different sizes. Bigger models handle harder problems but cost more. AWM reads the complexity of each task and routes it to the right model. Simple tasks go to faster, cheaper options. Complex ones go to heavier models such as Anthropic's Claude Opus or OpenAI's offerings. Pricing is fixed per task regardless, so companies can budget predictably without worrying that a run of difficult tasks will exhaust a credit allowance before the month is out.

What does this look like in practice?

At CoreWeave, a cloud computing company, product managers used to fill out detailed forms by hand whenever they launched a new product. Human reviewers would read those forms and manually assign tasks to finance, hardware, and marketing teams.

Under AWM, a product manager writes a standard document. An AI reads it, builds the project structure, and assigns tasks. One agent monitors overall progress and flags delays. Another forecasts infrastructure costs and flags when numbers fall outside historical norms. People review the AI's outputs and make the calls that matter.

The busywork moves to the machine. Judgment stays with the people who understand the business.

Should workers worry about being replaced?

Bose framed AWM as a tool that handles routine coordination so people can focus on decisions that require real judgment. The CoreWeave example supports that reading: humans still approve costs and evaluate recommendations.

The tension worth watching is that Anthropic and OpenAI, which supply the underlying models powering AWM, are also building their own agent products that compete directly with Asana. Bose acknowledged it. His argument is that 18 years of workflow data and pre-built processes for specific industries give AWM something a raw AI model can't quickly replicate. That's a reasonable claim for now, though it's exactly the kind of moat that gets tested fast once the big labs decide they want the enterprise coordination market for themselves.

Watch for: If your company starts using an AI agent connected to shared company data, ask your IT team how access controls work. A well-built system will limit what the agent can tell you based on what you're authorised to see. If nobody can answer that question clearly, the guardrails may not be in place yet.