Work
Putting AI agents to work
I build software where AI agents do useful work alongside people.
Most of the effort goes into the engineering around the model: what an agent can see, what it is allowed to do, and how you know afterwards what it did. These are the problems I spend my time on.
- 01
Agents that do real work
Assistants that answer questions from live operational data and carry out tasks, working within the same permissions as the people they help.
- 02
One capability layer for people and agents
Each operation is defined once as a typed capability and shared by the application, the assistants and an MCP server, so an agent can never do something the system does not already know how to do.
- 03
Guardrails by design
Trust tiers that decide what an agent may do alone and what needs a person to confirm, with a ledger recording every action an agent takes.
- 04
Software that repairs itself
A pipeline that watches production errors, groups them, works out the root cause and opens a tested fix for a person to review.
- 05
Codebases built for agents
A living map of the system generated from the code and checked on every change, shared rules, and verification gates, so an agent starts each task from an accurate picture.
- 06
Many agents, one project
Coordination tooling that lets several agents work in parallel on the same codebase, each in its own lane, without colliding.
- 07
Language models on messy data
Embeddings and language-model normalization to match inconsistent records against a clean source of truth, with a person confirming the result.
- 08
Keeping what the AI learned
Capturing the decisions and context that build up during AI-assisted development, so that knowledge stays with the team rather than on one laptop.