Context Engineering, Not Model Smarts, Is Blocking Production Agents

Atlan’s experiments found that building an agent is easy, but giving it accurate business context—knowledge, norms, and expertise—consumed most of the effort and led to trust failures. Context engineering is the new bottleneck, not model capability.

Their shift to a “context layer” with shared skills and data graphs, managed like code, mirrors human team learning. However, they immediately hit dependency management nightmares, context drift, and ownership blur. Skills reading from each other broke downstream when any one evolved.

The talk frames context as IP and envisions a “GitHub for context” with lifecycle management. But the evidence is purely anecdotal from one company; the solution is nascent and introduces new governance, security, and scaling challenges.

For engineering leaders, the takeaway is that agent orchestration is no longer just about model selection. It demands a new infrastructure category—context versioning, testing, and deployment. Without it, multi-agent deployments will become brittle and untrustworthy.