The CAFE(S) framework: Improving AI agent effectiveness

Engineering Enablement by DX

0:00

-39:20

The CAFE(S) framework: Improving AI agent effectiveness

Brian Houck and his co-authors explore how the CAFE(S) framework helps engineering teams improve context quality, reduce costly AI failures, and make agents more effective.

Brian Houck

Oct 16, 2026

Listen and watch now on YouTube, Apple, and Spotify.

As AI agents become more capable, context quality is increasingly the limiting factor in their performance. In this episode of Engineering Enablement, I’m joined by my CAFE(S) co-authors Max Kanat-Alexander (Capital One), Eirini Kalliamvakou (DX), and Margaret-Anne Storey (University of Victoria) to introduce our framework for evaluating whether agents have the context they need to do their jobs well.

We introduce the five dimensions of CAFE(S): Clarity, Actionability, Fidelity, Efficiency, and Security. We walk through common context failures, how poor context accrues technical and cognitive debt, what it costs in developer productivity and operating spend, how AI is changing the skills engineers need, and why context review could become as important as code review.

Some takeaways:

Why context quality matters for AI agents

Improving models and harnesses isn’t enough to maximize AI agent effectiveness. The quality of the context agents receive is just as important, yet organizations rarely evaluate it systematically.
Poor context creates compounding costs and risks. Ambiguous instructions, irrelevant information, and insecure context can lead to runaway token usage, incorrect outputs, security breaches, and additional developer toil.

The five dimensions of the CAFE(S) framework

CAFE(S) defines five properties of high-quality context: clarity, actionability, fidelity, efficiency, and security. Each addresses a distinct failure mode and helps identify where context needs improvement.
Clarity may be the most important dimension because ambiguity is often invisible to the person providing instructions. Agents can’t reliably infer unstated assumptions, and an incorrect interpretation can compound throughout a task.

Why model failures are often context failures

Agents frequently make mistakes because critical knowledge exists only in developers’ heads. Without documented requirements, constraints, and decisions, agents may build unwanted features or take actions that conflict with the team’s intent.
Security failures can stem from giving agents access to information they shouldn’t have. One real-world example involved an agent finding an unrelated API credential and deleting a production database.

How poor context contributes to technical, cognitive, and intent debt

Undocumented goals, constraints, and decisions create intent debt that can become more costly with AI agents. Agents need access to the reasoning behind decisions, including what a system should not do.
Better context can help prevent debt, but successful agent workflows can also encourage complacency. Teams may stop documenting important decisions when agents appear to be working well.

Creating high-quality context is becoming a core engineering skill

AI fluency involves learning to anticipate what agents need rather than correcting mistakes after they happen. More experienced developers proactively define goals, constraints, and success criteria based on lessons from previous failures.
Skills traditionally associated with senior engineers may become essential at every level. Clearly communicating intent and requirements has long been part of senior engineering roles, but AI agents are making those skills increasingly important for all developers.

Why context review could become as important as code review

Shared context deserves more scrutiny because its effects multiply across agent interactions. A mistake in a widely used instructions file can be repeated in every session, while an improvement can make every agent more effective.
Context review should be proportional to the importance and reach of the information. Critical shared instructions may need formal review, while lightweight documentation updates should remain easy to make.
Reviewing context can also build shared understanding across engineering teams. Like code review, the process can help developers align on requirements, decisions, and the intended behavior of their systems.