Agentic AI · 2026
Specflow
Engineering workflow · 7 teams · Architect and programme lead
A spec-driven development workflow where the specification is the artefact both humans and models work from — adopted across seven teams and thirty-five engineers.
reduction in development cycle time
teams adopted
engineers on the workflow
The problem
AI assistance was making individual engineers faster and the organisation no faster. Work arrived as tickets, context lived in people's heads, and generated code failed review for reasons no one had written down.
What I did
- Defined a specification format covering intent, contracts, failure modes and acceptance criteria, kept in the repo beside the code.
- Built tooling that generates implementation scaffolding, tests and review checklists from the spec.
- Rolled out team by team with a baseline measurement first, so the 60% cycle-time claim came from before-and-after data.
- Ran a weekly clinic for the first eight weeks — adoption failed everywhere I skipped it.
Architecture
A CLI and GitHub App over the model gateway, specs as versioned Markdown with a JSON schema, generation and review hooks in CI, and a metrics pipeline into the engineering dashboard built on Datadog, OpenSearch and Jira.
Request path
Similar problem on your side?