Kenny Brown

A scheduled loop where an AI agent independently pulls tickets from the team's tracker, investigates root cause, plans a fix, implements it, and opens a draft PR with generated test scenarios — never auto-merging. Governed by a token budget, clock windows, a kill switch, and an anti-rework ledger that caps retries per ticket to stop runaway loops.

Splits feature and bug work across specialized agents — research, test-writing, implementation, review, spec-compliance, QA generation — each seeing only what it needs. Implementation agents are kept blind to acceptance criteria (they see only failing tests) to prevent "testing to the answer," while cheaper models handle review and stronger models handle code generation.

After a database fix shipped on inspection alone and recurred in production within two weeks, I built a skill that forces any AI-authored stored-procedure change to compile and run against a real local database before it can be marked done — directly eliminating a recurring defect class caused by unverified LLM-asserted fixes.

A skill that interrogates a proposed plan adversarially until every open question and edge case is resolved. A variant cross-checks the plan against the project's domain vocabulary and architecture-decision records, updating them live as ambiguity clears. Packaged for reuse across projects.

Used an LLM to plan and partially execute large modernization efforts — a framework version upgrade, a front-end architecture migration, and data-access-layer changes — generating multi-week staged execution plans plus automated functional-parity audits that compare old vs. new component behavior before cutover.

Authored internal docs and a library of slash-commands and prompts — ticket planning, implementation, backend and frontend test generation — so other developers and QA staff could adopt the same AI-assisted TDD workflow, including setup for connecting an MCP server to the team's work-tracking system.

Set up a local open-weight coding model to handle low-stakes draft text (documentation, retrospectives) while reserving cloud-hosted frontier models for code generation and reasoning-heavy work — concrete cost-aware model selection inside a production workflow.