AI Agent Developer: Claude Code / Codex Agents That Fix Code and Open Pull Requests

We're building an AI-driven delivery pipeline for an engineering team: agents that take a piece of work from a report all the way to a tested, reviewed and deployed change. You'll start on the core, coding agents that run on their own, and then build out the rest of the pipeline. MAIN FOCUS: headless coding agents - Agents that run on a schedule or on events (a failing test, a new bug report, an error in the logs, a dependency update) - The agent reproduces the problem, writes a failing test first, makes the fix and opens a pull request with proof: tests failing before and passing after, plus screenshots - A second, separate agent reviews the first agent's tests and code, because an author shouldn't grade its own work - Setting up the agent's environment: sandboxes, scoped permissions, no access to production data - Logging what every agent run did, so any change can be traced and explained WHAT COMES NEXT - Intake: turning rough bug reports and feature requests into clear specs that point at the relevant code - Risk gates: rating each change by risk (database migrations, auth, shared data, test coverage). Low risk goes to an agent; high risk goes to a senior developer first. - Automated review: AI code review, dependency and package checks, security scans and before/after artifacts on every PR - QA evidence: Playwright runs with video, screenshots and traces, and an agent that checks the results against the acceptance criteria - Release and production watch: smoke tests after deploy, plus agents watching telemetry for errors, slow queries and regressions, with rollback decisions left to people - Learning loop: when a bug slips through, an agent works out why the earlier checks missed it and suggests improvements - Later: rebuilding data pipelines on a workflow orchestrator, with old and new outputs compared before each switch YOU SHOULD HAVE - Hands-on experience running AI coding agents (Claude Code, Codex or similar) outside a chat window: from scripts, on a schedule or in CI - Solid TypeScript, Python or Go - Comfort with Git, GitHub and pull request workflows - The ability to pick up an unfamiliar tool fast and get something working NICE TO HAVE - MCP servers, agent hooks or custom skills - GitHub Actions and required checks - Docker and sandboxed environments - Playwright or other browser automation - OpenTelemetry - Issue tracker APIs (Jira or similar) - Workflow orchestrators (Kestra, Temporal, Airflow) - PostgreSQL HOW THE WORK IS DONE - Everything lives in the repo as plain code, with short runbooks so the team can run it without you - Customer personal data never goes to AI tools ENGAGEMENT - Hourly, ongoing, 25–30 hrs/week to start - Remote, overlapping US Eastern hours until 5 PM ET