The context layer: why smart models still fail at work

I watched a talk by Prukalpa Sankar, founder and co-CEO of Atlan, that puts a name on something everyone building agents eventually runs into. Model intelligence has grown exponentially, yet most business AI projects still fail. Her diagnosis: the models are smart enough, but they lack the specific, situated knowledge a human expert carries. What is missing is not a better model. It is a context layer. These are my notes, plus what it looks like from where I build.

Performance is intelligence plus context

Pure cognitive ability only explains a fraction of human job performance. The rest is learned on the job: the facts, playbooks, and norms specific to one business. The same equation holds for agents. A frontier model dropped into your company with no context is a brilliant hire permanently stuck on day one. It can reason about anything and knows nothing that matters.

Humans shadow experts; agents get built in isolation

New people learn by shadowing someone senior, making mistakes, and getting feedback. Agents today are mostly built the opposite way: each team bootstraps its own, in isolation, with its own copy of the company baked into prompts. Sankar calls the result context sprawl. Ten agents, ten slightly different versions of the truth, and no way to tell which one is right. Atlan itself moved through this: they started with single-purpose agents and ended up rebuilding around one general-purpose agent backed by a shared company brain.

The context layer is GitHub for context

The framing I liked most: treat business context the way engineers treat code. That implies three properties.

• Lifecycle management: business skills and playbooks get versioned, tested, and collaborated on, not pasted into prompts and forgotten
• Compounding learning: agent traces feed back into the context library, so every run makes the next one slightly better
• Portability: context stays decoupled from any single agent framework, through open interfaces like MCP or formats like Iceberg, so the knowledge is never locked into one platform

In an era where models are increasingly commoditized, context is your intellectual property.

What this looks like from where I sit

This matches what I keep finding while building agent tooling and skills. The model is the interchangeable part; the versioned, tested instructions you wrap around it are the asset. A skill file that encodes how your company actually does an investor update or a code review is exactly the unit Sankar is describing: small, reviewable, portable context. The teams that treat that library as seriously as their codebase are the ones whose agents will feel like experts instead of interns.