Public-source agent readiness · 72 hours
Evidence you can show.Boundaries you can defend.
A fixed-scope engineering review for public AI-agent projects—mapping evidence, inference, tool authority, human approval, and the proof gaps between them.
Evidence / action map
Source-grounded reviewWhat the review resolves
Agents rarely fail at the demo. They fail at the boundary.
A persuasive answer is not the same as a reviewable system. AgentProof separates what the code shows, what the project declares, what a reviewer can infer, and what still needs runtime proof.
From source to claim to action.
See where a conclusion originated, how the model transformed it, which tool acted, and whether the evidence actually survives the handoff.
One explicit map of who—or what—can change state.
Up to 15 planned or executed scenarios. Every result carries its real status—never a made-up pass.
Ranked by launch impact and implementation effort, with code-level references where public source supports them.
Public sample · Incident Ledger
A report that refuses to pretend.
The sample review uses only the public repository, demo, architecture notes, and video. Dynamic checks that were not run are labeled Not verified—even where corresponding code or tests exist.
Strong prototype controls.
Insufficient production proof.
- Strongest control
- Source-enforced human authority boundary
- Largest gap
- Authentication, tenancy, and strict durable audit
- Runtime tests
- Not run in the public-source sample
Production identity boundary
RequiredStrict durable-audit policy
HighCommit-bound runtime evidence
HighHow it works
Three inputs. One bounded decision.
Send public links
One repository, one demo, and the primary workflow your agent should complete.
Freeze the scope
We agree the reviewed ref, authority boundary, package, price, and 72-hour start time.
Receive the evidence pack
A cited report, verification matrix, five prioritized fixes, and one factual-correction round.
Fixed scope · visible boundaries
Buy a decision, not an open-ended audit.
One public repository, one demo, one primary workflow, up to eight planned scenarios, and five prioritized recommendations.
- 72-hour delivery after scope + payment
- No credentials or private repository
- One factual-correction round
Up to three related workflows, 15 planned or executed scenarios, architecture-risk notes, and code-level references where possible.
Boundaries matter here too
What AgentProof is—and is not.
Do you need private access?+
No. The review is designed around public source, a public demo, and a clearly named workflow. Do not send credentials, logs, customer data, or employer material.
Is this a penetration test or certification?+
No. It is an engineering readiness review—not a vulnerability assessment, compliance audit, legal opinion, formal assurance engagement, or production-safety guarantee.
What starts the 72-hour clock?+
Confirmed payment, working public links, an agreed commit or branch, and a primary workflow that fits the selected package.
Will every scenario be executed?+
Only when safe, public, and within scope. Unexecuted checks are labeled Not verified. A source declaration is never reported as a test pass.
Founding intake
Make your agent reviewable before someone else asks.
Send the repository URL, demo URL, primary workflow, and any action that must remain under human authority. Do not send private data.
Request scope by email↗Payment or escrow link is sent only after scope confirmation.