Ship agents you can sign off on.

Leitwerk tests your AI agents against your own standards, proves the result with signed evidence, and tells you when an approval no longer holds.

The Readiness Check is a free 45-minute conversation.

Example run. 4 scenarios, 12 runs each. One broken hard rule rejects.

The problem

Three questions block every agent release.

1 of 3

What does “good” mean for us?

Generic benchmarks don’t reflect your standards or your risk appetite.

2 of 3

Can we prove an agent meets it?

Manual reviews are slow and leave no reproducible evidence.

3 of 3

Is the approval still valid?

Models, prompts and tools change. Yesterday’s test may not hold today.

Leitwerk answers all three. With evidence.

How it works

Four steps to a decision you can sign.

  1. Step 1. Define

    Short interviews turn your standards into testable criteria. One rule per requirement, written as policy code.

    Requirement to ruleRego

    “Refunds above the agreed limit need a human.”

    package leitwerk.refunds
    
    deny contains "refund.limit" if {
      input.tool == "erp.create_refund"
      input.args.amount > data.limits.refund
    }
  2. Step 2. Test

    Testing agents built for your setup check quality, scope and safety. Every scenario runs many times and every trace is scored by deterministic rules.

    Scenarios x runsExample
    48 traces scored48 passed, no hard rule broken
  3. Step 3. Approve

    A reproducible decision rule shows what is ready. You sign off. Everything goes into a signed evidence bundle.

    Decision ruleExample
    Approve ifno hard rule broken and lower bound ≥ required
    Hard rules broken0
    Lower bound, 95%0.93
    Required0.90
    VerdictAPPROVED
    Sign-off by your accountable ownered25519
  4. Step 4. Retest

    When a model, prompt or tool changes, you see which agents need reapproval. Before it reaches production.

    Change detectedExample

    Model endpoint updated

    • Refund agentReapprove
    • Invoice triageStill valid
    • Supplier onboardingReapprove

What you get

Approved. Against what. Until when.

One dashboard shows which agents are approved, against which standards, and until when.

ApprovalsExample
AgentStandardStatusValid until
Refund agentCustomer service v3ApprovedValid until 2027-03-31
Invoice triageFinance controls v2ApprovedValid until 2027-02-28
Supplier onboardingProcurement scope v1RetestModel changed
HR policy assistantPeople data v1TestingRun 31 of 48
$ leitwerk verify refund-agent.bundle
signatureed25519valid
traces48recomputed
scores48match
verdictAPPROVEDmatch
  • Signed evidence behind every approval.

    Built for audit. Anyone can verify it on a second machine: signature, traces, scores and verdict are recomputed.

  • Faster releases.

    Less manual review, and a decision that is the same every time you run it.

On premise

Your data
stays yours.

EU-hosted or in your own environment. On premise, Leitwerk runs fully inside your walls and needs no outbound connection.

2

CPU-only containers. No GPU needed.

0

Outbound connections needed. It runs fully on premise.

MCP

Your agent talks to the gate, not to your real ERP, CRM or mail.

Design Partner Program 2027

One real workflow.
8 to 12 weeks.

A pilot on one real agent workflow, with the first companies that want approvals they can defend.

  • Success criteria agreed before we start.
  • Direct influence on the roadmap.
  • Preferred terms as one of the first partners.

Letter of intent

Non-binding

It starts with one page that names four things.

Workflow
the agent workflow we pilot
Sponsor
who owns it on your side
Success criteria
agreed before start
Timeline
8 to 12 weeks

AI Agent Readiness Check

Start with
45 minutes.

A free conversation about where your agents stand. We score six areas from 1 to 5. Within two working days you get a scorecard and three recommendations.

  • Definition of “good”
  • Testing and evidence
  • Approval and accountability
  • Data and access control
  • Change and retesting
  • Overview and reporting

We also look at what makes it urgent:

  • A blocked release
  • Customer or audit pressure
  • EU AI Act, NIS2 or DORA
  • An incident or near miss

Book a Readiness Check

Or tell us you want to join the design partner program.

I’m interested in

Prefer email? Write to hello@leitwerk.tech.

Thank you.

We will get back to you to find a time. If anything is urgent, write to hello@leitwerk.tech.