The Behavioral Layer maps how agentic AI systems behave, disclose, escalate, and fail, and who is accountable for it. It is a working reference for the people building and governing AI agents.
Curated by Joel Goldfoot. See how this site is made.
A public, version-controlled knowledge base on the behavioral, trust, and experience layer of agentic products: the layer that sits above raw model capability and decides whether an agent is something a person can actually rely on.
Capability tells you what an agent can do. The behavioral layer governs what it should do, how it signals what it is doing, and what happens when it is wrong. That includes behavioral contracts, guardrails, escalation paths, failure and repair, confidence and disclosure, and the trust scaffolding that holds the whole interaction together. This site collects the methods, evals, frameworks, models, and research that make that layer legible, and reads each one through a single question: what does this mean for the people who build agents, and for the people who decide to ship them?
Every resource note here cites a primary source and is written for two readers at once
(a builder read and an exec read). The governing standard is EDITORIAL.md in the
repository root: accuracy is the brand.
Start here
- Behavior — the thesis core: contracts, guardrails, escalation, failure and repair, confidence, disclosure, trust scaffolding.
- Evaluate — evals, benchmarks, observability, and red-teaming for agent behavior.
- Build — agent frameworks, orchestration, memory, tool use, harnesses.
- Models — frontier models read through the agent-behavior lens.
- Research — papers, where the synthesis is the value.
- Signal — dated “what shipped / what changed” entries.
- Briefings — the exec lens and weekly digest.