eatp-reference

EATP is an open standard for establishing and verifying trust in enterprise AI agent systems. It separates trust establishment (human judgment, done once) from trust verification (machine-speed, continuous), tracing every action back to human authority through cryptographic chains.

Install with pip install kailash, then follow the four-operation lifecycle: ESTABLISH creates an agent trust chain, VERIFY checks permissions before actions, DELEGATE transfers constrained authority, and AUDIT records actions in a tamper-evident trail.

OAuth/OIDC handles user authentication and SPIFFE/SPIRE handles service identity, but neither governs agent autonomy. EATP adds agent trust delegation, constraint envelopes, and unbroken action-to-human traceability chains.

The postures are Pseudo-Agent, Supervised, Shared Planning, Continuous Insight, and Delegated, representing graduated autonomy. Postures upgrade through demonstrated performance and downgrade instantly when conditions change, including emergency downgrade to Pseudo-Agent.

BudgetTracker uses integer microdollars with a two-phase reserve/record lifecycle and fail-closed semantics. Threshold callbacks fire at 80%, 95%, and exhaustion, and SQLiteBudgetStore provides crash-safe persistence with 0o600 file permissions.

Key pitfalls include generate_keypair returning private key first, signatures being base64 not hex, constraints living on CapabilityAttestation rather than GenesisRecord, and reasoning signature verification requiring the parent record's context_id.