At VB Transform 2026, Asana’s Chief Product Officer Arnab Bose told VentureBeat’s Sam Witteveen how his team addressed a common enterprise AI shortcoming: chatbots that answer individual prompts but don’t remember prior interactions or whether past runs actually improved outcomes. The result is Agentic Work Management (AWM), an operating system that treats AI agents as coachable, multiplayer teammates rather than one‑to‑one assistants.
The limit of stateless chatbots
Most AI copilots today are stateless and tied to a single user prompt. They can fetch data from external tools but the execution is isolated—one‑off outputs that don’t become reusable workflows for the next person. Bose described how such calls fail to leverage a shared company brain or context graph, leaving future users without continuity.
Built on an 18‑year Work Graph
AWM is built on Asana’s 18‑year‑old architecture, the Work Graph. This graph database organizes information using what the company calls the Pyramid of Clarity: the smallest unit is a task with an assignee and due date; tasks roll up to projects, projects to portfolios, and portfolios to company‑level goals. The structure can trace, for example, how a delayed design task affects a revenue target. Because AWM plugs into this shared ledger, its agents can view company goals, update project statuses, and share memory with human colleagues rather than operating from a single user’s local context.
Access controls for confidential work
Deploying AWM for enterprises required solving data governance and boundary problems. If an AI teammate learns from workflows and human feedback, it builds shared memory—so the system must ensure that memory from a confidential project doesn’t leak to unauthorized employees who later interact with the same agent. Bose pointed to the example of a secret M&A project: someone who lacks access shouldn’t be able to leverage memories created for that project. Asana implemented access controls that distinguish when an interaction creates memory versus when it merely executes a task.
Dynamic model routing and hidden prompt engineering
AWM also abstracts prompt engineering away from users via dynamic model routing. When a user assigns a task to an AI teammate—say, drafting a job description—the system cross‑references public job postings, Asana’s internal style guide, and product requirement documents. For complex tasks, it routes the job to a heavier frontier model (Bose mentioned Anthropic’s Opus and OpenAI’s models as examples); lighter tasks are handled by faster, cheaper models. The goal is to make assigning a task to an AI feel as natural as assigning it to a human.
Predictable billing: a static cost per task
Dynamic routing creates billing unpredictability because different models and task complexities burn credits at different rates. To avoid customers having to reason about variable consumption or throttling their workers, Asana designed a pricing model that charges a static cost per task completion. The platform absorbs model selection, token counts, and run limits so enterprises get predictable pricing.
Persistent state and metadata recording
AWM addresses the statelessness problem by creating permanent state: when an AI teammate completes a task, the system records metadata about the completion—whether it improved project status and how it affected higher‑level goals. That persistent record builds a reusable company memory and supports successive users and workflows.
CoreWeave example: automating product launches
Cloud provider CoreWeave is an early adopter. Previously, product managers completed complex forms that human reviewers parsed into tasks for finance, marketing, and hardware teams. Under AWM, a product manager writes a standard Google doc referencing product requirement documents; a deterministic AI workflow reads the doc, creates the project structure, and assigns tasks. Specialized agents then execute: one watches overall project status and flags bottlenecks, another forecasts infrastructure costs and recommends approvals when numbers align with historical budgets. The system automates busywork while humans focus on evaluating AI outputs.
The frenemy problem with frontier providers
Bose acknowledged a market tension: frontier model providers such as Anthropic and OpenAI, whose models can power AWM, are also shipping competing agent products (for example, Anthropic’s Claude integrated into Slack as Tag). He said this is the reality companies must live with. Asana’s argument for AWM’s durability rests on 18 years of user experience and workflow data plus prebuilt standard operating procedures for industries—expertise that raw frontier models do not inherently possess. A lightweight model demo in Slack can be useful, Bose noted, but it typically requires curated channels and separate credentials for downstream apps, whereas AWM aims for true end‑to‑end enterprise use.
Conclusion
Agentic Work Management repurposes Asana’s Work Graph to create access‑controlled, persistent AI teammates that can update project state, share memory, and hide the complexity of model selection and billing from end users. The system is already in production with multiple customers, and Asana positions AWM as a solution to the real‑world bottlenecks of stateless enterprise agents.
