A trading platform built with agents, inside strict architecture tests.
Coding agents work within architecture rules enforced in CI, and a fleet of browser agents checks the live platform around the clock.
BlazorArchitecture testsAgent QA
Agentic engineering for software teams that already ship.
Your developers already have AI coding tools. This programme changes how your team specifies, builds, tests and reviews work around them, on your own codebase, so the speed shows up in delivery and not just in pull requests.
Run by Dom Roberts, who builds production systems with coding agents every day.
6 to 10 weeksalongside your developers
Your codebasereal backlog, not a demo project
Measureddelivery metrics before and after
The problem
Why the tools alone haven't made you faster.
Almost every developer now uses AI. Far fewer teams can show it in their delivery numbers. The research points to the same cause: the tools speed up typing, while the work around the typing stays the same.
90%
of software professionals now use AI at work, and over 80% say it has made them more productive.
Google DORA, State of AI-assisted Software Development 202519%
slower: experienced developers in a 2025 randomised trial took longer with AI, while believing it had sped them up.
METR trial, summarised in this review of the evidence+91%
time spent in code review on teams with high AI adoption. More pull requests merged, but bigger ones, and the queue moves to the reviewers.
Faros AI study of 10,000+ developers, in the same reviewDORA's own summary is that AI amplifies what is already there. A team with clear specs, good tests and fast review gets faster. A team without them gets more code to untangle.
The programme
What we change, week by week.
Every step happens on your repositories and your backlog, with your developers doing the work and me alongside them.
Week 1
Baseline
Measure where you are today from your own repositories and pipelines: lead time, review time, change failure rate, rework. Interview the team about where their time really goes.
Weeks 1 to 2
Context agents can follow
Write the context files, conventions and architecture notes that coding agents read before they touch your code, and agree a task format that gives an agent enough to work from.
Weeks 2 to 4
Guardrails in CI
Add the checks that catch bad agent output before a person has to: architecture tests, characterisation tests around legacy code, property-based tests and review gates in the pipeline.
Weeks 3 to 8
New ways of working, on real tickets
Pair with developers on live backlog items. Smaller pull requests, agent-first chores such as test backfill, upgrades and migrations, and a review checklist that keeps reviewers from becoming the bottleneck.
Final weeks
Measure and hand over
Re-run the baseline, write up the before and after, and leave a playbook in your repository with two or three internal champions who own it.
What you get
Left behind in your repository, not in a slide deck.
- A before-and-after delivery reportMeasured from your own pipelines, not a survey
- Agent context files and conventionsFor Claude Code, GitHub Copilot or both
- Guardrail tests and CI gatesArchitecture, characterisation and property-based tests
- A team playbookHow tasks are written, run, reviewed and merged
- Trained internal championsTwo or three developers who carry it on
- Data and security rulesWhat code and data each tool may see, agreed with you
Fit
Who this is for.
A good fit if
- You have between 5 and 200 developers shipping a product people pay for
- Your team already has AI tools but can't show the gain
- You work in .NET, TypeScript, Python, Java or similar, with a codebase that has history
- Leadership wants evidence before rolling it out further
Probably not yet if
- You have no automated tests at all and no appetite to add them
- You want a one-day training course and a certificate
- The decision to adopt AI hasn't been made. Start with Where AI fits instead
In practice
The same methods, in my own production work.
Rules written in Python, shipped in C#, kept in step by tests.
Conformance vectors generated from the prototype run against the production code on every build, so agents can port logic safely.
Conformance vectorsCI.NET
Questions
Agentic engineering, answered.
What is agentic engineering?
It's a way of building software where AI coding agents carry out whole tasks, such as writing a feature with its tests or upgrading a library, while developers write the specification, set the guardrails and review the result. The skill moves from typing code to directing and checking it.
Which tools do you use?
Whatever suits your team and your data rules. I use Claude Code and GitHub Copilot daily, and have worked with other agents and open-weight models. The practices carry across tools, so you aren't locked in.
Does it work on old, legacy codebases?
Yes, and that's often where it pays most. We start by wrapping the risky areas in characterisation tests, which lets agents make changes that a person can trust.
Will our code be used to train AI models?
Not if the tools are set up properly. We agree which plans and settings to use, which repositories each tool may see, and where sensitive work should run on models inside your own environment.
How do you measure the improvement?
From your own repositories and pipelines: lead time for changes, review time, change failure rate and rework, compared before and after. No self-reported surveys.
Related
Before you decide
Talk to Dom
Want your team building with agents?
Tell me what's prompting the question. I'll reply within one working day to arrange a short first call.
30mA call about your situation, not a sales pitch
1dA written note of what I'd look at first
£0No cost until we agree a scope