Systems
I build agent harnesses for companies and for individuals, on hosted models or on open-weight models running on their own hardware. Many of them are pointed at legal work, for law firms, in-house teams and legal-tech founders.
Agent harnesses
An agent harness is the scaffolding around a language model that makes it dependable for real work. I build it in six parts, and each build uses the parts the work needs.
Tools and MCP integrations
A model does useful work only through the tools it can reach. I write MCP servers for the tools people use every day, such as document stores, mailboxes, registries and case-management systems, and decide for each one what the model may read and what it may change.
Skills and workflows
A skill is a written procedure the model follows the same way every time: how to check a citation, or when to stop and ask a person. Workflows chain skills into a process with fixed steps, so that a run on Monday and a run on Friday can be compared.
Memory and context
The harness decides what the model sees at each step and what it keeps between runs: the documents of the matter, earlier decisions, a style guide. What it leaves out matters as much, because a crowded context gives worse answers and can carry material from one matter into another.
Multi-agent orchestration
Larger tasks are split across several agents. They run in parallel on separate sources, in a pipeline where one agent’s output is the next one’s input, or adversarially, with one agent checking or arguing against another’s work before anything reaches a person.
Evaluation and guardrails
Before a build is used, the client and I agree what working means for the task and write it down as test cases with expected results; every change is run against that set. Permission boundaries limit what the model can touch, and a person signs off before anything leaves the system.
Model backends
The same harness runs on a hosted model or on open-weight models on your own hardware, and the backend can be switched without rebuilding the rest. Where the material is sensitive, the default is local, as set out under private by design.
What you receive
Each build is made on demand and comes with:
- The software
- Documentation
- The evaluation set
- A licence
Legal capabilities
Most of what I build for legal teams is one of these harnesses, pointed at a legal task.
Research agents
A research agent takes a brief, runs it across EU and national sources, from legislation and case law to regulators’ publications, and returns a memo in which every proposition is cited. A lawyer reads the memo and checks the citations before anything is relied on.
Regulatory monitoring
A monitoring harness reads registries, official journals, filings and regulators’ notices on a schedule, picks out what touches the matters it has been given, and hands a short note to a named person, who decides what to do with it.
Contract review
Contract review runs a first pass over a defined set of clauses against your own playbook and marks where a draft departs from it and why. A lawyer checks every mark, and every clause outside the set.
Case-law models
Case-law models classify and rank decisions: by the issue decided, by the reasoning used, by how close they sit to the facts of your matter. Where the data allows, they give an estimate as a range, which the lawyer weighs as one input among others.
Blockchain contract tooling
I build smart-contract templates for recurring agreements, on-chain logic for settling defined disputes and credentials aligned with eIDAS. A lawyer drafts or approves the terms each contract encodes, and the code is reviewed before anything is deployed.
Private by design
I deliver local-first. Where the material is sensitive, the harness and the model run on your machines or on hardware you control, and the services around them are self-hosted. With local models, documents stay on your machines; with a hosted model, I name the provider and its data terms before anything is sent.
No client data flows to a third party unless you have agreed to it in writing.
The devices and accounts around a build count as much as the build itself, and I set them up for teams as part of my security work:
Evidence
Published work
One skill is public: pbl-socratic, which runs Socratic study sessions, on my GitHub profile.
Models I use
I build and test on Anthropic’s hosted API, including Claude Code and the Agent SDK, on OpenAI’s hosted API, and on open-weight models running on my own hardware. The harnesses switch between them, so a build is not tied to one vendor.
Licensing
Builds are licensed and maintained, not transferred. You receive the software, its documentation and the evaluation set it was tested against, and I keep maintaining it under the licence.
The licence fixes whose name the build carries and the purpose it is intended for, both decided per contract. Together they determine who is the provider and who is the deployer under the AI Act, so they are settled before the build starts.
Commissioning a build
Tell me what the work is, who checks it today and where the documents live. Most builds start with an assessment, described under Advisory.