APIR · Production HarnessProduction-ready, says the vendor who runs it.That’s not verification.
In July 2026, Microsoft’s VP of Core AI spelled out what a production agent requires. Every platform now solves that list inside its own walls, the auditor and the audited are the same company. APIR is the independent layer. We don’t run the agent. We prove what it did.
The Harness Check runs per agent: make a free account (or sign in) and you land on your fleet to pick one.
The bar just moved
Microsoft just told every enterprise what a production agent requires.
Its own identity, as a first-class principal
An audit trail that survives compliance review
Fleet observability with drift detection
Guardrails at the tool boundary
Continuous rubric-based evaluation
That’s the definition from Microsoft’s VP of Core AI, July 2026, and by their own public statements, 80,000 enterprises build on their platform. Every one of them just got the same bar to clear.
Here’s the part nobody says out loud: every platform vendor solves this list inside its own walls. The company that runs your agent also grades your agent. That’s homework, self-marked. APIR is the independent layer: the registry, the record, and the verdicts live outside the platform that executes the agent, where anyone can check them.
APIR is independent: not affiliated with, or endorsed by, Microsoft.
The pain → the proofSeven ways agents fail review.
Seven ways agents fail review.
Seven receipts that survive it.
01
“The AI did it.”
Anonymous actions. The audit trail collapses at exactly the moment you need a name.
Agent identity
Every agent gets a registry entry, a signed mandate, and a named accountable operator. An Ed25519-signed Trust Passport with a public verify page anyone can open. When something happens, there is a principal on record: not a shrug.
02
“Nothing we log survives compliance review.”
Screenshots and app logs are not a record. The reviewer knows it, and so do you.
Agent Black Box
Every action lands in a hash-chained, Merkle-anchored, tamper-evident record. Gaps get detected, not papered over. Publicly verifiable without an APIR login. Built against EU AI Act Article 12 record-keeping.
03
“Our metrics say it works. They don’t say it works right.”
Latency and uptime measure the plumbing. They say nothing about the decisions.
Behavioral Rubrics
Use-case-tied yes/no checks, scored continuously against the agent’s real Black Box record, not a synthetic benchmark. Rubric verdicts sit beside the Trust Score. They are never the Trust Score, and never blended into it.
04
“The model updated and the agent changed. Nobody re-ran the evals.”
The harness answer is to re-tune and re-run evaluations before shipping. Does yours?
Model-Change Watch
Any model change auto-opens a re-evaluation event, visible on the agent’s public passport until it’s resolved. Your buyer sees the open flag before you’d like them to: which is exactly why it works.
05
“The prototype shipped. Production readiness was somebody’s TODO.”
The demo worked, so it went live. Then it met a real customer.
Pre-Deployment Gates
A go/no-go verdict before an agent touches production. Pass the gate or don’t ship. The verdict is recorded, so “who signed off” is never a mystery later.
06
“A regression at fleet level is invisible until a customer finds it.”
One agent drifting is a bug. Ten drifting the same way is a pattern nobody was watching.
Fleet analytics
Rubric pass rates and open re-evaluations across the whole fleet, in one view. Honest empty states: where there’s no evidence, we show nothing: never a made-up number.
07
“A tool response carried injected instructions, and the agent followed them.”
The guardrail has to sit at the tool boundary, where the untrusted content enters.
Tool-boundary checks
Rubric checks at the tool boundary, plus Black Box evidence that screening controls existed and actually fired. APIR verifies the controls ran. APIR is not the runtime: and never claims to be.
Same principle, our layer
Their retrieval stack returns a structured “I don’t know” instead of hallucinating an answer. We hold ourselves to the same rule. APIR returns null scores when no behavioral evidence exists, and rubric verdicts include insufficient_evidence. An honest unknown beats a confident wrong answer. Every time.
The artifact
Walk into the review holding the file.
An independently verifiable artifact you hand to the auditor, the buyer, or the insurer: signed, sealed, and checkable without taking our word for any of it.
Sealed Evidence Pack
Per agent, on the rate card
One agent’s full record: identity, mandate, tamper-evident action chain, rubric verdicts: sealed into a single verifiable file.
Audit-in-a-Box
Full fleet, priced at checkout
The whole fleet in one artifact: every agent, every chain, every open re-evaluation: built for the reviewer who asks for everything.
Start where it costs nothing.
Check any agent free and see what an independent check finds. No signup, no card.
Check any agent free