A credit agreement in. A restructuring analysis out, every finding traced to its clause.
Agentman builds agentic harnesses with law firms and finance teams: specialist agents that read the documents, check each other's work, and write the first-pass analysis to your firm's standard. One is live with a global law firm on debt restructuring. Your partners review it and decide.
- Every finding carries its clause and page
- Agents check other agents before a lawyer reads a word
- Your methodology, held as versioned skills you can read
Live with a global law firm, on debt restructuring.
A debt-restructuring harness built with the firm analyses live deals today. Every figure below was counted from the live systems on 17 September 2026.
Mean score of the harness's decks against golden-standard decks written by the firm's partners.
* Mean of the four formally benchmarked deals.
The 43 agents, by job: 17 extraction · 4 forensics · 9 authoring · 4 adjudication · 2 grounding · 2 routing · 2 deal facts · 3 parse and voice.
- 10live deals analysed, US and European, in four currencies
- 69legal documents read, more than 4,300 pages
- 43specialist agents, each with one job
- 96versioned methodology skills, read at run time
- 4,740automated tests
Attributed to a global law firm. Figures as of 17 September 2026, counted from live systems. * 91.3 is the mean of the four deals formally benchmarked against partner-authored golden-standard decks. We do not name the firm, its matters or its clients.
From a 300–800 page credit agreement to a 12–18 slide analysis.
The first pass on a restructuring is structured work: find the terms, test them, and set out what they allow. The harness does that pass, and hands a partner an analysis with its sources attached.
- 01
Read
Extraction agents work through the credit agreement and the documents around it and pull out the terms the analysis turns on, each with the clause and page it came from.
- 02
Check
Before anything is written up, other agents test what was extracted: is the quote on the page, does the number match the clause, do two readings disagree.
- 03
Author
Authoring agents write the slides to the firm’s methodology and in its voice: the structure a partner expects, with the citation beside every finding.
What the analysis covers
Debt capacity
How much more debt the documents allow, under which baskets, and on what conditions.
Lender protections
The covenants, guarantees, security and consent rights that protect each lender group, and where they give way.
Enforcement mechanics
Who can accelerate or enforce, on what trigger, with what majority, and in what order.
Liability management pathways
The routes the documents leave open to restructure the debt, and what each one requires.
What goes in, and the slide a partner gets back.
Three slides from one fictional matter. On the left, what the harness was given; on the right, what it wrote, with the clause beside each finding and the open points flagged for the partner.
Example Holdings B.V.: first-pass restructuring analysis for the lender group. Start with debt capacity.
- Senior Facilities Agreement612 pagesPDF
- Intercreditor Agreement148 pagesPDF
Debt capacity
Baskets available
- General basket: Greater of €75m and 20% of EBITDA · cl. 24.3(b), p. 318
- Ratio debt: Unlimited below 4.5x senior secured leverage · cl. 24.3(c), p. 319
- Incremental facility: €120m, same security · cl. 2.4, p. 41
For the partner
- EBITDA add-backs: Uncapped synergies in the definition (p. 22) enlarge every basket above
- Slide 6 of 14. Draft for partner review.
Illustrative · fictional matters
How much more debt the documents allow
A fictional senior facilities agreement goes in. The slide that comes out lists each basket the borrower could use, the amount, and the clause it rests on. Where two definitions pull against each other, the slide says so instead of picking one.
Agents check other agents.
No agent's output is taken on trust. The agents that read and write are followed by agents whose only job is to test what they produced.
Forensics
When a number is disputed, forensics agents trace it back to the clause it came from and show the path.
Fact judge
Rules on whether a stated fact is supported by the documents. Unsupported facts do not go forward.
Evaluation judge
Scores the finished analysis against the standard the firm’s partners set.
Source verifier
Checks that text quoted from a contract is on the page it is cited to.
The document reader made up contract text. We caught it and changed the model.
During the build, the agent that reads the documents was caught fabricating contract text: sentences that read like the agreement and were not in it. In legal work that is the failure that matters most, because it looks right.
We moved the document reader to a different model. It is also why the checks above exist as separate agents: a quote has to be found on the page it cites, and a number has to trace to its clause.
Your firm's standard, written down as versioned skills.
The way your partners analyse a deal is captured as skills: written procedures the agents read at run time. The firm's standard lives in those skills, where a lawyer can read it and change it.
Read at run time
An agent reads the current version of a skill each time it runs. No analysis is hard-coded into the system.
Versioned
A change to the methodology is a new version of a skill, so there is a record of what the standard was on any run.
Written in your terms
A skill is a written procedure a lawyer can read and correct: how the firm tests a basket, what it calls a gap, how a slide is laid out.
Scope, build, operate.
A harness is a custom co-build with your partners and your innovation team, one workflow at a time. There is no list price: each engagement is scoped to the work.
- 01
Scope
We pick one workflow with you: the documents that go in, the deliverable that comes out, and examples of that deliverable your partners consider right.
- 02
Build
We capture your methodology as versioned skills, build the specialist agents and the agents that check them, and score the output against your partner-authored examples.
- 03
Operate
The harness runs on live matters. We run it with you, and tighten the skills and the checks as your lawyers review what it produces.
What is included
- The specialist agents for your workflow, and the agents that check them
- Your methodology captured as versioned skills
- Scoring against golden-standard work your partners write
- Operation and improvement once it is live
Ownership
Ownership and productization rights are agreed per engagement, before the build starts.
Security
Our SOC 2 and ISO 27001 audits are underway, and neither has issued a report yet. Before any document is shared, we walk your security team through how it is handled. Security at Agentman
Bring one workflow and an example of the finished work. We will scope it with you.
What partners and innovation leads ask
What is live today?
A debt-restructuring harness, built with a global law firm. As of 17 September 2026 it had analysed 10 live deals, US and European, in four currencies, reading 69 legal documents and more than 4,300 pages. We do not name the firm, its matters or its clients.
What does the harness produce from a credit agreement?
A 12–18 slide restructuring analysis from a 300–800 page credit agreement. It covers debt capacity, lender protections, enforcement mechanics and liability management pathways, with the clause and page beside each finding. It is a first pass for a partner to review, not a finished opinion.
How does it learn the way our firm does the analysis?
Your methodology is captured as versioned skills: written procedures the agents read at run time. The live harness reads 96 of them, and no analysis is hard-coded. When your standard changes, the skill changes and the next run follows it.
How do you stop it inventing contract text?
Agents check other agents. Forensics agents trace a disputed number to its clause, a fact judge rules on whether a fact is supported, a source verifier checks quoted text against the cited page, and an evaluation judge scores the finished analysis. We learned why this matters the hard way: during the build the document reader was caught fabricating plausible contract text, and we moved it to a different model.
How good is the output?
The decks score a mean of 91.3 out of 100 against golden-standard decks written by the firm’s partners. That is the mean of the four formally benchmarked deals, as of 17 September 2026.
Is it only for debt restructuring?
Debt restructuring is what is live. A harness is built for one workflow at a time, so M&A, finance and other document-heavy analysis are scoped as their own engagements. We will tell you plainly if a workflow is not a fit.
What does an engagement cost, and who owns the result?
There is no list price. Each engagement is scoped to the workflow, then built and operated with your team. Ownership and productization rights are agreed per engagement, before the build starts.
What is your security position?
Our SOC 2 and ISO 27001 audits are underway, and neither has issued a report yet. Before any document is shared, we walk your security team through how it is handled.
Which analysis should your partners stop starting from a blank page?
Tell us the documents that go in and the deliverable that comes out. We will tell you plainly whether a harness fits.
Live with a global law firm · Figures as of 17 September 2026