Hector encodes the full propose–validate–integrate cycle as a machine-speed protocol. Here's how the agent team, the core loop, and the seven infrastructure primitives fit together.
Hector ships with seven built-in agent roles, each with a defined scope, capability set, and system prompt template. New roles can be added as JSON files in .hector/roles/.
Owns the brief. Takes the initial input, decomposes it into structured requirements, manages the backlog, and communicates with the human supervisor.
Decomposes complex work into independently executable subtasks, each sized to complete within token and timeout budgets.
Writes, tests, and documents code. Works in a Claude Code session with file access, iterating with tools before submitting a structured proposal.
Reviews proposals against acceptance criteria and hard gates — failing tests, type errors, missing coverage. Returns a typed verdict: accept, reject with feedback, or escalate.
Owns information architecture, UI/UX decisions, component specifications, and design tokens. Produces structured artefacts consumed by the Engineer.
Manages build pipelines, infrastructure provisioning, monitoring, and deployment. Handles the "ship" part of "build and ship."
After each task, captures what worked and what didn't into episodic memory. Distils pattern-level insights to feed into future context assembly.
.hector/roles/The orchestrator drives this loop continuously, picking the highest-priority ready task, assembling context, invoking the right agent, and processing the result.
The PM agent takes the brief. The Planner agent breaks complex work into independently executable subtasks, each with acceptance criteria and typed dependencies.
The Engineer agent receives a tailored context payload — relevant files, decisions, and artefacts — and works with full tool access: read, write, build, test.
The QA agent receives the proposal and checks it against acceptance criteria and hard gates. It returns a structured verdict with specific, actionable feedback.
Rejected proposals go back to the Engineer with QA's feedback. Up to three revision rounds are allowed. On the fourth failure, the task escalates to a human supervisor.
Accepted proposals are committed to the Git-backed artefact store. Downstream dependencies are unblocked. The Retrospective agent captures what happened into episodic memory.
These are the building blocks that make autonomous agent collaboration reliable. Each solves a specific coordination problem that existing tools leave to you.