Across every company studied, one layer stands out as universally unsolved: context management. Not the prompt. Not the model. The organizational knowledge that an agent needs to do useful work.
This is not just an engineering problem. When you start building agents for a real company — say, a local services business with a hundred-thousand-dollar contract — the first thing you realize is that the agent needs to understand the organization. What does the company do? What are the standard operating procedures? Where is the institutional knowledge stored? The ops person is trying to onboard you, and the agent needs to be onboarded too.
The harder version of this problem is that organizational context is contested. One person thinks the company strategy is X. The CEO thinks it is Y. A truly useful context layer would not just store information — it would surface disagreements among humans and provide a way to resolve them. That is a product that does not exist yet.
Most companies are trying to solve this with whatever they already have — SharePoint, Confluence, Notion, Slack channels. None of these were designed to be a context layer for agents. But the right approach is probably not to force people onto new tools. The philosophy should be: use their existing tools, integrate with what they already have, and make the context layer flexible enough to sit on top of whatever data stores they are already using.
This is where Glean-style organizational search meets agent infrastructure. But Glean is search. What agents need is something closer to a living, queryable, version-controlled knowledge base that multiple agents can read from and write to — and that surfaces conflicts rather than hiding them. I've argued elsewhere that the build-vs-buy data shows context is the most universally homegrown layer in the stack — which is exactly why the company that finally cracks it captures the next floor of the platform.
Sources
Related Essays
Organizational Context Is the Hardest Problem Nobody Has Solved
Context management is the layer most consistently built in-house and least well served by vendors. It is not a search problem. It is a knowledge management problem.
What the Build-vs-Buy Data Actually Shows
From Stripe to a five-person startup, the agent stack is mostly blue — built in-house. The harness is bought. The middle of the stack is built. The opportunity sits in turning blue dots green.
Each Person Needs Their Own Agent Instance
Shared agents with a single configuration are fundamentally broken. Each user gets a personal copy that learns through interactions — and an organizational layer that locks in what consistently works.
Key takeaways
- Context management is the one layer every agent team is forced to build themselves. No vendor has cracked it yet.
- Organizational context is contested — different stakeholders disagree on facts. A useful context layer surfaces conflicts rather than hiding them.
- The right architecture sits on top of existing tools — Confluence, Slack, SharePoint — rather than asking the org to migrate.
FAQ
Why is context management harder than the model or the harness?
Because the model and harness are technical problems with technical solutions, while context is a sociotechnical problem. It requires capturing what the company actually does, who decides what, and how decisions get reversed. None of that is encoded in any single system of record.
Is this just enterprise search with a new label?
No. Enterprise search retrieves documents. A context layer for agents needs to be queryable, version-controlled, multi-writer, and capable of surfacing disagreements between humans. It is closer to a living knowledge graph than to a search box.