The thesis core. This section covers the behavioral layer itself: the rules, signals, and recovery paths that decide whether an agent is trustworthy in use, independent of how capable the underlying model is.

What belongs here

  • Behavioral contracts: the explicit promises an agent makes about what it will and will not do.
  • Guardrails: the constraints that keep behavior inside the contract.
  • Escalation: when and how an agent hands control back to a human or a stronger process.
  • Failure and repair: how an agent detects it is wrong, discloses it, and recovers.
  • Confidence and disclosure: how an agent communicates certainty, limits, and provenance.
  • Trust scaffolding: the patterns (previews, confirmations, undo, transparency) that let a person extend or withdraw trust safely.

What goes elsewhere

  • How you measure these behaviors lives in Evaluate.
  • The mechanisms that implement them (orchestration, memory, tool use) live in Build.
  • The evidence base (papers) lives in Research.