Asana's AI agents share organizational memory
Agents collaborate with humans to strengthen corporate memory
1The big hurdle of enterprise AI
Companies developing AI agents face a fundamental problem: chatbots can respond to a single prompt but can't remember previous interactions or evaluate past versions' performance. Asana's Chief Product Officer Arnab Bose explained their solution, Agentic Work Management (AWM), at VB Transform 2026. The system positions AI agents as collaborative teammates that learn alongside humans rather than one-off assistants.
218 years of data architecture
Asana built AWM atop their 18-year-old Work Graph architecture. Their Pyramid of Clarity structure connects tasks (smallest units) to projects, projects to portfolios, and portfolios to company objectives. For example, the graph can trace how a delayed design task impacts revenue goals. Work Graph provides a real-time ledger showing who does what, when, and why across the organization.
3Multiplayer teammate
AWM differs from standard AI assistants by leveraging Work Graph. While traditional copilots are single-user and temporary, AWM can view company-wide goals, update project statuses, and share memory with human colleagues. As Bose noted, the system works from a shared organizational ledger rather than individual prompts or local files. Early adopters like FedEx have reported successful implementations with measurable benefits.
4Security for confidential work
Deploying AWM to enterprises required solving critical data governance challenges. The system addresses boundary problems: when an executive creates workflows for confidential projects (e.g., M&A), the AI's updated memory won't leak context to unauthorized employees interacting with the same agent later. Asana implemented granular access controls determining what creates persistent memory versus executing one-off tasks.
5Dynamic model routing
AWM abstracts prompt engineering from users. Complex tasks (e.g., drafting executive job descriptions) route automatically to heavy frontier models (like Anthropic's Opus or OpenAI's offerings), while simpler tasks use faster, cheaper models. Users assign work naturally, as they would to human colleagues. Bose emphasized that knowledge workers shouldn't need to engineer perfect prompts or attachments for every task.
6Predictable billing
Since agentic tasks vary in computational complexity, credit burn rates could become unpredictable. To prevent customers from throttling AI usage due to cost uncertainty, Asana designed a static cost-per-completed-task model. The platform absorbs variability in model selection, token counts, and run limits, providing enterprises with consistent, forecastable pricing.
7The stateless chatbot problem
Current enterprise AI deployments often use stateless chatbots. While Model Context Protocol (MCP) integrations connect LLMs to tools like Slack or Google Drive, these agents lack persistence. For example, a marketing campaign draft pulls relevant data but remains a one-off execution. The workflow isn't reusable for the next person creating a similar campaign, creating redundant effort across teams.
8Creating permanent state
AWM establishes permanence by recording task metadata. The system tracks how completions improve project statuses and advance higher-level goals. Early adopters like CoreWeave are overhauling product launch processes using AWM. This approach aims to boost organizational efficiency by creating reproducible workflows that compound institutional knowledge rather than reinventing processes repeatedly.
Original article: VentureBeat — Asana'nın yapay zeka ajanları şirket hafızasını paylaşıyor