Hands typing on a laptop keyboard in low light

← All services

AI-assisted engineering

Agents, skills and hooks are how I build every day, with human review on everything that ships. I can set that up for your team.

What it covers

  • Skills and hooks and project rules that keep AI-written code reviewable and consistent.
  • Pull-request watchers, monitors and status helpers that take the routine out of release work.
  • Release gates: nothing merges without review and passing checks.
  • Coaching non-engineers who write code with AI help, with ownership files and rules that keep them safe.
  • Local models for work where data must stay in-house, benchmarked on your tasks.
  • A worked example: fleet-watch runs a small model locally on its hub to give a first read of each security incident, with the client choosing where the hub runs and the evidence staying where they decide. It suggests; a person decides.

Agents with real access, safely

An AI agent only does useful work when it has real access to code, servers and data. So the access has to be narrow, short-lived, logged and approved by a person. In July, AI models under evaluation at OpenAI escaped a light sandbox through an unpublished flaw in a package-registry proxy and reached Hugging Face's systems. After that I went well beyond a light sandbox for my own agents:

Egress through an allow-list

Agents reach the network only through a proxy with a list of permitted destinations. Everything else is refused and logged.

SSH certificates per task

Access to a server is a certificate minted for one task, restricted to one source address and short-lived.

SQL through a controlled integration

Database queries go through a controlled integration with its own limits. The agent never holds a database password.

A secrets runner

A command that needs a secret gets it injected at run time. The value is scrubbed from everything the agent can read.

A guard the agent cannot edit

The rules live in a guard owned by root, outside the agent's reach. It cannot rewrite its own limits.

A person approves

Risky actions stop at a deterministic approval step: the same request always asks the same question, and a person answers it.

The background is on the Security page.

How it runs

  1. Look at how you build today

    Where review is slow, where code lands unreviewed.

  2. Set the rules in the repository

    Skills, hooks and owners files, versioned like code.

  3. Automate the routine, keep the judgement

    Agents prepare; people decide.

  4. Check what changed

    Review time, defects and surprises, before and after.

A good fit when

Your developers already use AI assistants, and you want the speed without the unreviewed risk.

Hands-on training: your developers at AI speed, with the guardrails in place.

Tell me how your team works today; I will tell you what I would set up first.

Get in touch