Architecture

A trust layer that sits around agent execution.

AgentProof does not replace an agent framework. It wraps the run boundary, captures the causal sequence, enforces deterministic policy, and emits proof that can be replayed later.

Run lifecycle

01 / Intent
02 / Evidence
03 / Tool request
04 / Policy decision
05 / Replay
06 / Proof report

Boundary model

The server owns policy decisions and persistence. The client renders approved trace data, scrubbed evidence, and report exports. Reasoning explanations improve readability, but deterministic policy remains the source of execution authority.

intentplanevidencetool_requestpolicydecisionreportintent

Agent adapter

Normalizes intent, evidence, plan, and tool request events from any agent runtime.

Policy guard

Applies deterministic rules before a tool request can leave the trusted boundary.

Persistent trace store

Stores replayable runs in a file-backed JSONL path today with a sqlite upgrade path for teams.

Proof exporter

Generates redacted public reports and structured JSON proof bundles.