Mandates and evidence for AI agents in client work

Your AI agents did the work.Now prove it to your client.

Large buyers now write AI rules into supplier contracts: disclose the AI, log every agent action, approve before an agent starts, stop it at once. Axynera applies those rules while Claude Code works, and gives your client a signed evidence pack they can check offline, clause by clause.

Claude Code supportedRuns on your machinesVerified offline by the client
P1 m-6d204eb0 ACME-412 running

Move the appointments API to v2

issued by CTO → claude-code@lantern · approved by Dana (Acme security)

$5.04 spent · $4.00 held for a sub-agent · $8.96 free · of $18.00

10:42Edit src/appointments/api_v2.py · logged
10:43Bash pytest tests/appointments -q · logged
10:44Edit infra/prod.tf · blocked: outside the mandate
Evidence pack for Acme Health Signature valid
metClient approved before the agent started
metAgent stayed within its files and budget
not metAll AI disclosed to the client before use

Illustration. Lantern Ops and Acme Health are fictional.

$0.46metered by Axynera, matched the Anthropic Console to the cent
0.49 sfrom the client pressing Stop to the agent being cut off
0agent actions before the client approved the work
2 of 2tampered evidence packs rejected by offline verification

Live tests with real Claude Code and the real Anthropic API, 31 Oct and 2 Nov 2026.

Try it, no sign-up

See it from both sides in two minutes

01 · demo consoleThe supplier and the client viewA supplier runs Claude Code on a client's project. Switch View as to see what the client approves, stops and checks.Open the demo → 02 · live viewWatch agents work, press StopWhat the client sees while the supplier's agents act on its code, replayed from a recorded run.Watch the replay → 03 · clause scannerPaste your client's AI clauseSee which requirements can be shown with evidence and which cannot. The text never leaves your browser.Scan a clause →

The problem

Your clients' contracts changed. Your evidence didn't.

Agents now write code, run commands and open pull requests on client projects. Clients have noticed: AI clauses are appearing in the supplier terms of large healthcare, consulting and consumer-goods buyers. They ask what AI you use, who approved each agent, what it did, and whether they can stop it.

Today's answer, 1

A questionnaire

A yes in a spreadsheet says nothing about what an agent did on the client's code last Tuesday.

Today's answer, 2

Logs on your side

Your client cannot check them, and has no reason to trust a record the supplier can edit.

Today's answer, 3

A person checks every change

Reviewing each agent action by hand cancels the speed you sold the client in the first place.

Clause by clause

What the client asks for, and what Axynera records

The client asksAxynera records
Disclose the AI you use Every model and provider seen by the gateway and by the agent's own telemetry, checked against what you disclosed. An undisclosed model shows as not met.
Log agent actions and tool calls Each tool call with its time, a short summary and SHA-256 hashes of input and output, in a hash-chained log. Secrets are masked before anything is written.
Approve before an agent starts The agent's mandate stays pending. Model calls and file edits are refused until a person on the client's side signs it.
Stop an agent at once A client stop blocks the agent's model calls within a second, including a call already running. Only the client can resume the work.
Notice before a new AI provider or sub-agent An agent can hire a sub-agent only on narrower terms, and the client gets a notice with its task, budget and files before it starts.

Requirements paraphrased from public supplier terms. Not legal advice: your contract decides what counts.

How it works

Before, during and after every agent task

01 · Before

The client approves the work

Each agent task becomes a signed mandate: what to do and why, the files it may change, its budget and how autonomous it is. Nothing runs until a client approver signs.

m-6d204eb0 · pending → approved by Dana

02 · During

Your engineers work as usual

Claude Code runs through the Axynera gateway and hooks. Every model call is metered, every tool call logged, edits outside the mandate are blocked, and the client can stop an agent mid-task.

Edit src/appointments/api_v2.py · sha256 4f1c…

03 · After

The client gets proof

You sign an evidence pack that maps each of the client's requirements to what was recorded. The client checks it offline and signs its answer: accepted, or questions.

verify-pack pack.json → VALID

The evidence pack

One file your client can check offline

Signed by you, verified by themIt carries every mandate, signature and client approval. Verification runs offline, with no access to your systems.
Four honest statesMet, partly met, not met, not observed. Each requirement says what Axynera can see and what it cannot.
Tamper-evidentChange one value after you signed, a budget or a status, and verification fails.
What it does not proveWork done outside Axynera. A signature shows the record was not changed after signing; anything Axynera could not see is reported as not observed.
The client's answer on recordThe client signs its review into the same pack, so the question and the answer stay together.
Patient portal modernisation pack-a95fc389

Lantern Ops → Acme Health · signed by CTO · key ceb5d6f69faddffe

4 met 1 partly met 2 not met
metCapabilities, autonomy and scope documented
partly metActions logged and kept 12 monthsLogging recorded; retention can be shown after 12 months.
metClient approved before the agent started
metClient can stop the agent at once1 stop · 0 calls let through
metAgents stayed within scope
not metAll AI disclosed before usegpt-x-mini (openai) used, never disclosed
not metNotice before a new AI provider
Client review · signed Questions. “gpt-x-mini was used but never disclosed to us. Please explain.” — Dana, Acme security

Illustration with fictional companies.

Who it is for

Three people, one record they can all check

Supplier

Delivery head

Answer the client's AI clause with evidence instead of a promise. One pack per engagement, ready before the client asks.

Client

Security and vendor risk

Approve agent work, stop it when something looks wrong, and check what happened against your own AI clauses.

Supplier

Engineers

Keep using Claude Code. Axynera sits in front of the model API and in Claude Code's hooks, so there is no new tool to learn.

Questions

What it does, and what it does not

Which agents does it work with?

Claude Code today, through its API and its hooks. The gateway also speaks the OpenAI API (Codex CLI and others); those agents are not yet tested live. We add agents with pilot customers.

Where does our data go?

In a pilot Axynera runs on your machines. Your client receives only the evidence pack you choose to send. Secrets are masked before anything is logged.

Is this a certification?

No. It is evidence for the named clauses of one client's contract. It does not replace SOC 2, ISO 42001 or a legal review.

What can it not see?

Work done outside the gateway and hooks, such as AI in a browser chat or another editor. The pack reports that as not observed, never as met.

Could a supplier fake the record?

In a pilot the log runs on the supplier's machines. Signatures and the hash chain show that nothing changed after signing; they cannot prove work Axynera never saw. If your client needs more, we set up client-held keys and a copy of the log on the client's side during the pilot.

How does a pilot work?

One client engagement over a few weeks. We map that client's AI clauses, set Axynera up with your engineers, and produce the first evidence pack together. Scope and price are agreed on a short call.

Early access

Start with one client that asks about your AI.

We are choosing the first IT service firms to pilot with. Tell us which client and which clause, or paste it into the clause scanner, and we will show you what the evidence would look like.

Request a pilot sales@axynera.com