What Agent Proof Runtime is
Agent Proof Runtime (APR) is a proof runtime for autonomous AI work. Its job is not to make an agent sound more confident. Its job is to turn an execution into an evidence trail that can be checked after the fact.
The runtime is built around a simple engineering question: can another verifier determine what ran, what passed, what changed and whether the evidence was altered? Useful output alone does not answer that question.
APR deliberately separates what the agent produced from what the system can prove about the execution.
Proof before trust
Mission Manifest
The mission starts with explicit expectations. The manifest defines the task boundary and the checks that matter. That gives the runtime something concrete to validate instead of treating an agent response as success by default.
Deterministic checks
APR runs checks whose result can be evaluated independently of the model's prose. The distinction matters: a model saying “done” is a claim; a deterministic check is evidence about whether a defined condition passed.
Event chain and Merkle root
Execution evidence is chained and summarized so later mutation is detectable. The design is intended to make tampering visible to the verifier, not to pretend that local software provides hardware-backed trust.
Proof Bundle
The output of the proof boundary is a versioned evidence package. The verifier can re-check hashes, events, the Merkle root and bundle integrity without accepting the agent's own statement as the source of truth.
Honest verification labels
APR does not claim hardware-backed trust. The public runtime uses explicit labels so verification strength is not overstated.
| Status | Meaning |
|---|---|
| LOCAL_VERIFIED | The Proof Bundle passed the local independent verification path defined by APR. |
| FAILED | One or more required checks or proof validations failed. |
| UNANCHORED | The proof is not presented as externally anchored or hardware-backed. |
| development-only | The path is explicitly marked as development scope rather than production trust infrastructure. |
The useful property is not “trust the runtime because it says verified.” It is “inspect the evidence and run the verification path again.”
What sits inside APR
- Mission Control — a visual control surface for running and inspecting proof-oriented missions.
- Proof Bundle — the versioned evidence package produced by an execution.
- Independent verifier — re-checks evidence-chain integrity and proof contents.
- Tamper Lab — mutates a copy of evidence so detection can be demonstrated instead of merely asserted.
- Mission Studio — a multi-agent pipeline operating behind one proof boundary.
The repository also includes documentation for controlled live validation and hosted deployment validation. Public claims should be read together with those artifacts, not as marketing-only assertions.
Logs are useful. Proof is a different requirement.
Observability answers questions such as “what happened in the system?” Proof-oriented execution adds stricter questions: which mission boundary applied, which checks passed, which evidence belongs to this run, and can an independent verifier detect a changed bundle?
This distinction becomes more important as AI agents receive tool access, write files, call APIs, operate across multiple steps and act with less direct human supervision.