Scale expert work. Keep expert control.
An agentic harness turns one of your expert workflows into a working system: specialist agents, your method written as versioned skills, and checks on every output. Your experts review the result and make the call.
- Live with a global law firm on debt-restructuring analysis
- Your standards as versioned skills the agents read at run time
- Agents check other agents before an expert sees the work
A debt-restructuring harness, live with a global law firm.
A 300 to 800 page credit agreement goes in. A 12 to 18 slide restructuring analysis comes out, covering debt capacity, lender protections, enforcement mechanics and liability management pathways. The firm’s method lives in skills the agents read at run time. No analysis is hard-coded.
*Mean of the four formally benchmarked deals. All figures as of 17 September 2026, counted from live systems.
- 43specialist agents
- 96versioned methodology skills, read at run time
- 10live deals analysed, US and European, in four currencies
- 69legal documents, 4,300+ pages
- 4,740automated tests
One workflow. Scoped, built and run with your experts.
We start with a single workflow and a clear definition of a good result, and expand when the evidence supports it.
- 01
Scope
We pick one workflow with your experts: what goes in, what must come out, who reviews it, and what a good result looks like. You leave with a written scope.
- 02
Build
We capture your method as versioned skills, build the specialist agents and the checks around them, and test on representative cases until your experts accept the output.
- 03
Operate
The harness runs on real work with agreed oversight. We keep it running, your team reviews results, and the skills are revised as your method changes.
There is no price list. Scope, what is included and commercial terms are agreed for each engagement, as are ownership and productization rights.
A working system built around your standards.
Not a chatbot and not a prompt library. A harness is everything a group of agents needs to do one job to your standard, and everything your team needs to supervise it.
Specialist agents for your workflow
Each agent has one job: extract, check, write, judge. Small jobs are easier to test and to trust than one agent asked to do everything.
Your method as versioned skills
Your standards are written down as skills the agents read at run time. No analysis is hard-coded, so a change in method is a new version of a skill, not a rebuild.
Connections to the actual work
The agreed documents and systems are connected, so the harness works from the source material your experts use.
A test set and a quality bar
Representative cases, automated tests and a benchmark your experts set. Quality is something your team can measure, release after release.
Review points and a record
Agreed points where a person decides, and a record of what each agent did and which source it relied on.
Operation after launch
We run the harness with you, look at the hard cases, and improve the skills and checks as the work changes.
What goes in, and what your experts get back.
Three kinds of work on the same harness. The first is live with a law firm, the second is in build with a design partner, and the third runs today in Medman. The cases shown are invented.
First-pass restructuring analysis for the lender group. Flag anything that limits new money.
- Credit agreement612 pagesPDF
- Intercreditor agreement148 pagesPDF
Restructuring analysis
Debt capacity
- General debt basket: Capacity remains. Clause 7.01(k)
- Incremental facility: Ratio test not met at the last reported quarter
Lender protections
- Transfer of material assets: Restricted. Clause 7.05(d)
- Priming debt: Majority-lender consent only. For partner review
Next step
- Returned to the partner with two open questions.
Illustrative · fictional parties and patients
A credit agreement becomes a first-pass restructuring analysis
The harness reads the full agreement set, applies the firm’s method skill by skill, and drafts the analysis: debt capacity, lender protections, enforcement mechanics and liability management pathways. Each finding points to its clause.
The partnerReads the analysis, settles the open questions, and decides what the client is told.
No agent marks its own work.
In the live harness, the agents that write the analysis are not the agents that test it. Four kinds of check run before a partner opens the deck.
Forensics
Separate agents go back through the documents looking for what the first read missed or got wrong.
Judges
A fact judge tests the claims in the draft. An evaluation judge scores the finished work against the standard.
Source verifier
Checks that what the analysis cites is in the source document, where it says it is.
Evaluation agents at each stage
Every design and implementation is reviewed by evaluation agents at each stage, independent of the agent that did the work, before it lands.
Then a person decides. The harness prepares the work and shows its sources; your expert accepts it, corrects it or sends it back.
Review security and governanceStart where expert judgement meets repeated work.
Teams whose product is expert judgement
- Law firms, clinical programs and other practices where senior people review every deliverable
- Leaders who can state the standard a good result must meet
- Experts willing to grade early output so the harness learns the bar
A workflow that is a good first candidate
- It repeats, and each run takes days of careful reading
- It draws on long documents or records from several sources
- It ends in a deliverable an expert can accept, correct or reject
Running a medical practice? Medman is the first product built on the harness, live in independent practices.
Explore Medman43 specialist agents, each with one job.
This is how the debt-restructuring harness divides the work. Your harness is divided around your workflow; the principle is the same.
| Function | Agents | What they do |
|---|---|---|
| Extraction | 17 | Read the agreements and pull out the terms the analysis depends on. |
| Forensics | 4 | Re-read the documents for what the first pass missed or got wrong. |
| Authoring | 9 | Write each part of the restructuring analysis from the checked record. |
| Adjudication | 4 | Judge the facts and score the draft against the firm’s standard. |
| Grounding | 2 | Tie each statement back to the document it came from. |
| Routing | 2 | Send each document and question to the specialist that handles it. |
| Deal facts | 2 | Keep the facts of the deal consistent across the analysis. |
| Parse and voice | 3 | Turn long documents into readable text and keep the deck in the firm’s voice. |
| Total | 43 | Counted from the live system, 17 September 2026. |
Questions about an engagement.
What is an agentic harness?
A harness is the working system around a group of AI agents: the specialist agents themselves, the skills that hold your method, the connections to your documents and systems, and the checks that test the output before a person sees it. Agentman builds that system around one workflow your experts already do by hand.
What is running in production today?
A debt-restructuring harness is live with a global law firm: 43 specialist agents reading 96 versioned methodology skills, used on 10 live deals as of 17 September 2026. Medman, the first product built on the harness, is live in independent medical practices. A clinical case-preparation harness is in build with a design partner.
How does an engagement run?
In three stages. Scope: we agree one workflow, its inputs, its deliverable and the standard a good result must meet. Build: we capture your method as versioned skills, build the agents and checks, and test on representative cases with your experts. Operate: the harness runs on real work with agreed oversight, and we keep improving it with you.
What does it cost?
There is no published price. Each engagement is scoped to one workflow, and the scope sets what is included: the build, the connections, the test set, and operation after launch. Talk to us about your workflow and we will scope it with you.
How does the harness follow our standards?
Your standards are captured as versioned skills that the agents read at run time. No analysis is hard-coded. When your method changes, your team revises the skill, the change is tested, and the next run uses the new version.
How is the output checked?
Agents check other agents. Forensics agents re-read the documents, a fact judge tests the claims, an evaluation judge scores the finished work, and a source verifier checks the citations. Then your expert reviews the result and decides. The harness prepares the work; it does not sign it.
How do we know it works?
We agree acceptance criteria before the build and test against examples your experts have already done well. In the live debt-restructuring harness, decks score a mean of 91.3 out of 100 against partner-authored golden-standard decks, across the four deals that were formally benchmarked.
What do we need to bring?
A workflow that repeats, access to the people who know how it should be done, representative examples you are authorized to share, and a clear view of what an acceptable result looks like. We identify the systems and documents involved during scoping.
Who owns the resulting work?
Ownership and productization rights are agreed per engagement. The agreement separates your data and your firm-specific material from Agentman’s underlying platform and reusable components, and states who may do what with each.
Which models and deployment options can we use?
We scope the supported models and deployment options against your quality, cost, security and data-handling requirements. What is available depends on the workflow and how it is deployed.
Which expert workflow should scale next?
Tell us what needs to be delivered, how often the work repeats, which documents and systems it touches, and how your experts judge a good result.
Start with one workflow and a clear definition of success.