Control evidence for AI coding agents

Agents write the code. Your controls still apply.

What the agent changed. Who approved it. The evidence your auditor asks for.

Session recordedsession cc-7f2a · fincard/payments-apiDemo data
  1. 09:41:02session.startedclaude-code · payments-api
  2. 09:41:09intent.bundle.attachedspec PAY-2141 · sha256:a41f…
  3. 09:41:33tool.startedEdit src/ledger/settle.ts
  4. 09:41:34approval.requestedconfirm · payments-write
  5. 09:42:10approval.decidedapproved · second reviewer · rationale
  6. 09:43:05hook.blockedBash curl → public host
  7. 09:43:41model.completed1 secret redacted before emit
  8. 09:46:01session.finishedPR evidence digest sha256:f9b0…
Illustrative session · 8 events · 9 control references.

01Industry trend

Adoption outpaces governance.

2 Dec 2027EU AI Act rules for Annex III high-risk systems apply, including logging (Art. 12) and human oversight (Art. 14). European Commission, AI Act (opens in a new tab)

Sources: Stack Overflow Developer Survey 2026 (self-reported; data under ODbL 1.0) and the Veracode GenAI Code Security Report. AutoDevOps is not affiliated with either publisher.

02How it works

Questions your auditor will ask.

63 controls · 18 frameworks
  1. Who approved the agent’s change?

    The reviewer, the decision, and the rationale.

    Human approvalEUAI.ART14-03
  2. Which agent wrote this code?

    Agent, session, and commit, on the pull request.

    Change attributionFFIEC.SDLC-02
  3. Can the agent log be trusted?

    Signed events, append-only, with hashes you can check offline.

    Agent event recordsEUAI.ART12-02
  4. What if no reviewer is available?

    Gated actions are blocked, and the block is recorded.

    Approval outage handlingDORA.RES-03

Approval gates require the customer-cloud distribution. The public CLI records activity.

03The record

Every decision, on the record.

Sensitive actions wait for a named reviewer. Decisions export as hash-checked evidence, never raw prompts or source.

Illustrative portal view · demo data

Approval queue in the FinCard demo portal: a packaging command pending second-line review, a Bedrock rerun approved by a named reviewer, and a public model endpoint call blocked by the policy engine.

Confirm

A high-risk command waits for a reviewer.

Approve

Tied to the session, command, and reviewer.

Block

A public model call is denied and recorded.

You set the thresholds, own the review, and validate the deployment. Evidence is not a certification. See evidence exports

Connect in one command.

Supported today: Claude Code, Codex, and Cursor. Copilot CLI is roadmap. Setup guide

terminal
npm install -g @autodevops/verifier
verifier install --harness claude-code --mode telemetry

Prompt text is opt-in.

Built for your cloud.

Where agents run

  • Developer laptops

    Claude Code · Cursor · Codex hooks

  • CI pipelines

    Verifier hooks on pre-commit and pre-push

Amazon Web Services
Validated
Portal stack in your account
In progress
Azure and Google Cloud
Roadmap