An agentic workflow that ships production code
A way of working where coding agents do most of the typing, and the engineering goes into the specs, context and checks around them.
The problem
Agents write plausible code quickly. Speed stops being the limit and confidence takes its place: a team needs to know the work is correct before it ships, and reviewing everything by hand gives the time back.
The skill is engineering the context around the model.
Decisions
Every task starts from a written spec. Each project carries context files and reusable skills, so an agent begins with the conventions a new teammate would have to learn. Each ticket runs in its own git worktree with its own simulator and Metro port, which lets several agents work in parallel without colliding. Automated checks gate the output: lint, types and tests, plus custom checkers that block a result when it breaks a rule. Agents drive the simulator to verify UI changes, and nothing merges without review. The approach is model agnostic; at the time, the tools were Claude Code, Cursor and MCP servers.
Teaching it
I built and led The New Stack, a hands-on internal workshop for 25 engineers at Infinite Red on context engineering: how to give an agent context, how to give it tools, and how to break work down for it. Participants run the same task with and without a context file and compare the results.
Result
Agentic development is now part of how my team at Infinite Red works every day, and the workshop gives other engineers a structured way to start.