Inside an agentic software factory
- agentic-ai
- software-factory
- delivery
- public-sector
How a small public-sector team organised coding agents, a backlog and a set of gates into a factory, and what went wrong along the way.
· 12 min read
Service
Agentic software factory
Coding agents in your delivery process, with the specifications, tests and review gates that let you ship what they write.
At a glance
- Format
- Assessment, 30-day pilot, factory set-up
- Duration
- From a few days to several months
- Where
- Remote, on site in Luxembourg and the EU when it matters
- Languages
- English, French
When it helps
Sounds familiar?
- Your developers use coding agents, but releases are no faster: the work now waits in review and testing.
- The code agents write runs, but it duplicates logic, ignores your architecture and makes every next change harder.
- Your stories are too vague for an agent, which fills the gaps with plausible defaults.
- You want to start, but security, data location and cost need answers first.
What I do
What I take on
Intent and requirements
Stories with acceptance criteria that can fail and Gherkin examples, refined before anything is generated.
Factory set-up
Rules files, specialised subagents, commands and skills, committed in the repository so everyone works the same way.
Guardrails and gates
Hooks, separate agents for tests and code, contract checks and CI gates that an agent cannot get past.
Traceability
From story to scenario, test, code and merge request, with CI checking every link.
Hosting, data and cost
Models in an EU region where required, bounded access to code and secrets, and cost tracked per project.
Team practices
New roles and review habits, and one rule above the others: you must be able to explain what you merge.
Ecosystem
What a factory is made of
Agents do the repeatable production work. People own the intent, the judgement and the acceptance. These are the pieces that hold the two together.
Agent tooling
- Claude Code Coding agent
- MCP Agents connected to your tools
- Subagents One narrow job each
- Skills Know-how, packaged
- Hooks Rules enforced while it works
Engineering practices
- Gherkin scenarios Executable acceptance
- Test-first Tests written by another agent
- Contract-first API One OpenAPI spec, both sides
- Isolated worktrees One per story
- ADRs Decisions owned by people
The AI-native SDLC playbook from Anthropic describes six stages of an agent-assisted lifecycle, from plan to maintain. I use it as a shared map when a team starts.
Formats
Ways to work together
Assessment
A few days
Where agents fit in your delivery, what slows you down today, and the guardrails your context requires.
Pilot
30 days
One team, one product area, one lane of the board, starting from a starter harness of rules, agents and hooks that I adapt to your stack. We measure cycle time and rework, and add a gate each time something goes wrong.
Factory set-up
Several months
Alongside your team, from refinement to CI, until the factory runs without me.
- 2026
Agentic software factory · EU Publications Office
Rewrite of a large public web platform, organised as an agentic software factory: Claude Code configured in the repository with 16 subagents, 27 commands and 12 hooks, CI gates and a knowledge graph. From April to October 2026: more than 1,000 merge requests and 34,000 automated tests.
- 2019 – 2026
CTO · Agile Partner
Member of the company’s AI task force: chose the tools, defined good practices and a starter harness for coding agents, and helped colleagues build their skills.
Blog
Related writing
Other ways I can help
Contact
Let's build something good together.
Tell me about your context and timeline. I usually reply within one business day.