AI product engineering practice

AI-assisted development: agentic coding

A licence for every engineer doesn’t make a team faster. A playbook, review rules and honest measures do. Week 1 ends with a plan and a price, in writing.

A paper workbench where a jointed arm lowers a vermilion block into a wall of paper bricks, a magnifying glass checking the joint and a banded notebook beside it.

What the service covers

We start by reading how your team works today, what slows it and what blocks it. Then we pick the tools and patterns that fix those problems.

  • A tool choice, in writing

    We try coding agents on your real backlog and write down which one to standardise on, and why.

  • A prompt and context playbook

    Patterns for the recurring jobs in your stack, written down so your team reuses them and new engineers learn them fast.

  • A review policy for agent code

    Agent-written changes are labelled, checked against an extra list and sampled for deeper audits. Nothing ships unread.

  • Measures that resist gaming

    Lead time to production, change failure rate and time to restore, before and after. We don’t count lines of code.

3 ways to engage

Most teams start small and widen what measured well. Week 1 settles the scope, in writing.

  1. 1What you get

    Discovery week

    We read how your engineers work, what slows them and what blocks them, and name the tool and patterns to try.

    How it runs

    One week of discovery that ends with a plan and a price, in writing.

  2. 2What you get

    Pilot rollout

    One team, one tool, one playbook and one outcome you can measure, to build the case for wider use.

    How it runs

    Fixed scope and fixed cost, agreed before work begins, with a named person on every gate.

  3. 3What you get

    Organisation-wide rollout

    Several teams: tooling, training, review policy and measurement, ending with a playbook your organisation keeps up to date.

    How it runs

    Scaled only where the pilot measured well. Documented handover, or we keep operating it with you.

Who the service is for

AI-assisted development / Fit check

It fits

  • Engineering teams weighing coding agents, or partway through a rollout.
  • Engineering leads answerable for delivery who want their own numbers, not a vendor’s case study.
  • Security and platform leads worried about agent-written code reaching production.

It doesn’t fit

  • One engineer working alone. Our published notes are free to read.
  • Teams that have already standardised and want a stamp of approval. We will find the friction.
  • Teams that want agents to ship code nobody reads.

Questions about this service

Every engagement runs the same method: an agent proposes, a person approves.

Which coding agent should we standardise on?
It depends on your codebase, your privacy needs and how your engineers already work. In the pilot we try candidates on your real backlog and write down the trade-offs. In our own shop, Claude Code drafts code, tests and documentation under an engineer’s direction.
How do you keep agent-written code safe?
Agent-written changes are labelled in the pipeline and reviewed against an extra checklist: tests, new dependencies, hard-coded values. Dependency and secret scans run on every change, auth code needs a senior reviewer, and each sprint a sample of merged changes gets a deeper audit.
How do you measure the gain without gaming it?
We measure lead time from commit to production, change failure rate and time to restore, before the rollout and after it. We don’t count lines of code or commits, because those numbers reward the wrong habits.
What does it cost?
Fixed scope, fixed cost, agreed before work begins. We don’t publish a rate card because engagements differ; the price is in writing in week 1. When choosing a tool, fit matters more than its licence price.