Intake — A ticket enters the pipeline.
From Jira, the customer portal, or the API. AE inspects prerequisites, checks the repo, and opens a work item.
PRD — Ambiguous tickets become concrete asks.
The product agent refines the ticket into a PRD grounded in your codebase. You can review it before implementation begins.
Research — The codebase is read, not guessed at.
Research agents scan relevant files, identify patterns, and summarise what the implementation actually needs to know. No hallucinated findings — only what was actually found.
Planning — Work is decomposed into ordered subtasks.
The implementation planner produces code-focused tasks with dependencies. No doc tasks. No git-history tasks. Just the code.
Implementation — Code is written, tested, verified, then committed.
TDD by default. Tests verified before commit. Three retries with feedback on failure. Git operations are deterministic — the LLM never commits unilaterally.
Review — A separate agent reads the diff.
Approves or requests changes. Feedback loops back into implementation with three more attempts.
Human review — You see the actual diff.
Syntax-highlighted, in the approval queue. Approve, request changes, or reject. No summaries standing in for the code.
QA — The real test suite runs on the feature branch.
Not a mock. Not a subset. If tests fail, a bug loop routes the fix back to implementation with the failure context.
PR — A pull request is opened.
The description is derived from the real diff — never a hallucinated summary.