An AI harness: the controlled workplace an agent runs in
The agent does the repetitive work inside a harness: one work order, frozen inputs, a written mandate, your checks on every output, one run record and a named person signing every release.
| Inputs | hashed, versions pinned |
|---|---|
| Agent | inside a written mandate |
| Checks | yours, on every output |
| Record | one per run and rerun |
| Release | named person signs |
- 01Work orderaccepted whenthe deliverable, its tests and its reviewer are written down
- 02Inputs frozenaccepted whena rerun on the same inputs gives the same result
- 03Agent worksaccepted whenthe agent can use only the tools its mandate lists
- 04Checks decideaccepted whena failed check sends the work back
- 05Person signsaccepted whennothing is released before a named person signs
- 06Record keptaccepted wheninputs, steps, checks and review sit in one record
Parts
What a harness is made of
H01
Work order
One named deliverable with its inputs, acceptance tests and a named reviewer. The agent works on nothing else.
H02
Frozen inputs
Data, specifications and reference versions are hashed and pinned before the run, so a rerun gives the same result.
H03
Mandate and tools
A written list of what the agent may read and do, enforced by the tools it is given. No credentials it does not need.
H04
Checks
The checks your team already trusts run on every output: rules, validators, comparisons and reconciliations. A failed check sends the work back.
H05
Run record
Inputs, versions, every agent step, check results and the review, stored as one record per run and rerun.
H06
Named sign-off
Nothing is released until a named person reviews the findings and signs. The agent holds no release rights.
Division of labour
Judgement stays with people
People decide
- What the specification means where it is silent
- Whether a finding is a real defect or an explained exception
- Whether the output is fit to release, and they sign for it
The harness does
- Freezes inputs and pins versions
- Runs the agent inside its mandate and captures every step
- Runs the checks and blocks release until a named person signs
How we would start
Scoping
We map one workflow from input to sign-off, the checks you already run and the first deliverable for the pilot.
Parity on past work
The harness redoes completed work and we compare it with what your team delivered. Nothing touches live work.
Live, under review
The agent takes one seat in the live workflow, for example the independent check. People adjudicate every difference.
Enablement
Your team learns to build the next harness on the same controls.
Harnesses
Where we apply it
- AML harness
The agent gathers evidence and drafts the case file. The analyst decides.
Capability - Clinical research AI harness
Frozen inputs, pinned CDISC standards and a named programmer signing every release.
Capability - AI and governance
The controls every harness runs on: inventory, mandates, sign-off and an audit trail.
Controls
Common questions
Q01Is a harness a product?
No. It is the way we build AI into a regulated workflow. Each harness is built around your checks, your data and your sign-off rules, and you keep it.
Q02Which work fits a harness?
Work that already has objective checks: rules, validators, reconciliations or a second person comparing outputs. The checks are what make the agent's work safe to accept.
Q03Which models do you use?
The model is a replaceable part. The run record stores which tool, model and prompt version produced each output, as provenance and never as a verdict.
Q04How is it priced?
A fixed-fee scoping, then milestones against acceptance tests agreed in writing. No price before scope.
Next step
Which of your workflows already has the checks?
A scoping call with Oleg. We pick one workflow and plan a parity study on your own past work.
Oleg Pun
CEO, Vareger
