Work / reviewable-agentic-development

Agent engineeringDec 2025 – Sep 2026

Agent-assisted development as a reviewable practice

I made AI sessions part of code review, with a rubric, named red flags and an onboarding gate. Velocity without reviewability is debt.

Claude CodeCodexBMADGitHub ActionsNotion APITypeScript

The problem

A team that moves to agent-assisted development produces code nobody reviewed the thinking behind. The diff looks fine; the reasoning that produced it is in a chat window that closed. Velocity goes up and reviewability goes down, and the debt shows up months later as code that nobody on the team can defend.

What I did

I brought an agentic software lifecycle and Claude Code into the company on the first day of the engagement, then built the rails around them in three layers.

Tooling. Agent commands living in the repository, so the workflow is versioned with the code. A workflow that syncs stories between the product tracker and GitHub, so that agent-authored story files in git are the tracker of record and a human can read what an agent was asked to do. A written internal guide with an explicit rule for when not to use an agent.

People. An AI-coding course that new developers complete before they start work. A four-phase program that moved one engineer from supervised use to autonomous agentic development. And the rule that carries the whole practice: review of the AI session is part of the code review. Reviewers score prompting quality, review discipline and workflow efficiency out of five each, and look for named red flags: pasted output without review, hard-coded credentials, more than fifty messages spent on one task.

Orchestration. A work tracker in which each item advances through discover, session and resolve nodes, and where the agent doing the work writes its own confidence, risk and suggested next step instead of a heuristic computing them. Any node can run against one model, a council of models, or a head-to-head comparison. Beside it, an adversarial multi-model review tool: several rounds of structured disagreement between three frontier models in architect, red-team and synthesiser roles, with a judge. I ran nine such reviews on real design decisions.

Why it matters

The question a hiring manager should ask about any team using agents is not how fast they are but whether they can still defend their code. This practice answers it with an artifact: every reviewed change carries the session that produced it and a score on how it was produced. That is the seam between agents and people, made reviewable.