The agent pipeline
Every run walks through the same seven stages, in the same order.
Every PilotX run — however it was triggered — walks through the same fixed sequence:
Scout → Analyst → TestWriter → AutoCoder → QA Reviewer → Runner → Reporter
The order never changes and no stage is skipped, though a run can pause between QA Reviewer and Runner if your workspace requires human approval (see Human gate & approvals), and a run can stop early if a stage fails outright.
How it's orchestrated
Runs execute as durable Temporal workflows, not a single long-lived request — each stage is a discrete unit of work with its own recorded status (pending, running, succeeded, or failed) and timestamps, visible on the run's detail page as it happens.
Reasoning vs. execution
Scout, Analyst, TestWriter, AutoCoder, QA Reviewer, and Reporter each make a real call to an LLM (Anthropic's Claude) to do their actual reasoning — reading the ticket, scoping a plan, writing test cases, proposing code, judging quality, and writing the summary. If the model call fails or the stage isn't configured, each of these stages falls back to a deterministic version of its output rather than failing the run outright.
Runner is different: it doesn't call an LLM. It executes each eligible test case for real, inside an isolated, single-use sandbox that's created fresh and torn down after every run — see What each stage does for what "eligible" means.
Custom agents
Admins can insert custom agents between any two stages of this pipeline from Settings → Custom agents, and override the system instructions of the built-in agents from Settings → Agent instructions. The fixed seven-stage order above is what ships by default for every workspace.
Next
See what each of the seven stages actually does: What each stage does.