Practice area
The hard part was never getting an agent to write code.
It is making the output safe to merge — repeatably, in a regulated product, without a human reading every line and without pretending that a human did.
This practice is the substrate around the agents rather than the agents themselves: what each one may touch, how collisions are prevented, what review it must survive, and what record it leaves behind.
-
14,219
commits across 30 repositories in 9 months, by one engineer directing governed AI agents
-
50
skill catalogue entries, with persona-based governance manifests per agent role
-
90
parallel work lanes under file-claim leasing, so concurrent agents never collide
A commit count is a measure of throughput, not of quality. It is here because it is the honest size of the experiment; the 94% coverage figure on the governance practice is the one that says whether the throughput was worth having.
What I do here
Governance before autonomy
The interesting problem is not getting an agent to write code — it is making the output safe to merge. Persona-based manifests bound what each agent role may touch; claim leasing stops two agents editing the same file; every autonomous action lands in an audit trail.
Adversarial review as a gate
Review agents whose job is to attack the change, not approve it, sitting in the pipeline as a quality gate rather than a suggestion. Provenance hashing and anti-hallucination verification mean a claim in a pull request can be checked against something, instead of taken on trust.
Evaluations as infrastructure
If you cannot measure whether the system got better, you are not engineering it, you are hoping. Evaluations run like tests — versioned, in the pipeline, blocking — which is what turns a demo into a delivery capability.
On the record
2025 – 2026
A governed multi-agent delivery platform
Turned LLM coding agents into a production-grade delivery capability across a 30-repository codebase: persona-based governance manifests, a 50-skill catalogue, file-claim leasing so parallel agents never collide, adversarial review agents as quality gates, and audit trails on every autonomous action. Output included a greenfield twelve-service governance platform, an event-sourced IoT control plane, and shipping firmware.
2025 – 2026
One policy boundary for people and agents
The compliance authority layer governs humans and AI agents through the same rules. That is the property that makes autonomous action permissible in a regulated product: an agent is not trusted because it is capable, it is constrained because the boundary does not care what kind of actor is asking.
“He didn’t just support the shift from a traditional Scrum flow to AI-native development processes: he helped design it, rethinking how work gets scoped, estimated, reviewed, and shipped when AI tooling becomes part of the team.”
This is one of four practices.
It only works on top of the other two — a policy boundary that treats an agent like any other actor, and an architecture whose contracts are explicit enough to review.