Asana says agent fixes must transfer team-wide, or everyone trains a different AI

If you’ve ever taught an AI agent to do a task the “right” way, and then watched that skill vanish for the next person who uses it, you’ve already met the real enterprise problem. In VentureBeat reporting, the issue is blunt: when someone on a team corrects an AI agent, that improvement disappears the moment a colleague opens the same tool. The correction does not transfer, so the next person effectively starts from zero, rebuilding understanding from scratch.

Asana Chief Product Officer Arnab Bose frames the stakes in one sentence to VentureBeat: model providers are getting much better at improving reasoning and retry loops, but they are not good at bringing enterprise work context in a way that humans can reason about for shared memory. Bose says Asana built its agentic platform to solve exactly that. In Asana’s Agentic Work Management platform, if any team member corrects an agent, that correction applies to everyone else on the team, with a “context graph” automatically provided to agents operating inside Asana’s system. That means teams do not have to turn every human into an expert at prompt engineering or context engineering just to get consistent agent behavior.

Why this matters now is that the labor numbers are already telling you AI is moving from “experiment” to “workflow,” but the payoff is uneven. According to Asana’s own research, 75% of knowledge workers use AI on the job, but only 5% of companies have reported productivity gains. In other words: adoption is real, but so is the gap between using AI and getting repeatable improvements. Shared memory is the missing bridge when organizations expect one set of agent behaviors to hold across users, tasks, and days.

Bose also argues this is not only an Asana product question, but an enterprise design decision for any multi-agent system. The core problem is structural: the models powering agents are stateless by design, so memory has to live as a dedicated layer outside of a context window. While AI research and product teams are racing toward maturity in what to store and how agents reason, VentureBeat notes that key questions remain largely unsolved: what gets stored, who controls it, and how the memory stays consistent when different agents and users write to the same instance. That becomes unmanageable in the very scenario enterprises care about, where agents are supposed to work with the entire team.

This is where second-order failure modes show up. VentureBeat reports that many platforms still treat agents as individual-first, meaning agents act for individuals and learn user-specific preferences. The result is that task repeating becomes normal, and reality becomes inconsistent across the organization as each user’s version of the “truth” evolves. Agents could even contradict each other. In a multi-agent workflow, those contradictions are not just annoying. They can spread mistakes, because “the agent that learned from one person” may not align with “the agent that learned from another,” even if both are operating under the same umbrella tool.

Collate co-founder and CTO Sriharsha Chintalapani adds a more granular explanation in an email to VentureBeat. Agents are sensitive to prompt quality, and someone with a strong understanding of the task tends to produce more accurate results, partly because they can construct better prompts and partly because they can give the agent better feedback. Chintalapani’s key point is that the agent remembers the corrections it receives and applies that knowledge to successive prompts. But that “remembering” is only meaningful if it is shared. Organizations, he argues, should not treat shared memory as a prompt engineering problem. They should build systems that repeat context across every conversation.

Zeta Global’s chief data officer Neej Gore takes the compounding angle even further in a separate email: shared context becomes a living memory that compounds intelligence across the enterprise. The governance implications are real too. If memory is personal, the enterprise gets local wins and global maintenance. If memory is shared, the enterprise can build institutional knowledge automatically. Microsoft’s Copilot illustrates the individual-first approach by storing memories tied to a user’s role, tone preferences, and working patterns across Microsoft 365 surfaces. For teams evaluating agentic platforms, that’s not necessarily wrong. But for engineering and orchestration teams, VentureBeat reports that shared memory is now turning into a procurement criterion rather than a technical nicety. An agent that learns only for the person using it requires ongoing individual upkeep, while a connected team-wide memory layer can reduce repeating work and align behavior.

For boards, founders, and enterprise operators, the strategic takeaway is straightforward: the value of agentic systems is not just in better reasoning or better retry loops. It is in whether improvements can propagate across the organization without resetting every time a new user logs in. In an environment where companies already see widespread AI usage and still struggle to report productivity gains, shared memory is the operational difference between “agents that help individuals” and “agents that scale trust.”