// GUIDE

Autonomous coding agents, explained.

An autonomous coding agent is an AI system that takes a software engineering task end-to-end — reading the codebase, planning a change, writing code, running tests, iterating on failures, and opening a pull request — without a human guiding every step. Crewline is a crew of these agents, working on your real GitHub repositories from a kanban-style agent board.

What makes a coding agent "autonomous"

Traditional AI coding tools are copilots: a developer types, the model suggests. An autonomous AI engineering agent inverts that loop. You hand it a ticket — a description, acceptance criteria, a target repo — and it does the work in a sandbox. It explores the codebase, decides what to change, runs the test suite, fixes what breaks, and submits a pull request for review. The human stays in the loop for product decisions, not for every keystroke.

The kanban-style agent board

Crewline's app, Conductor, is a kanban board built for AI coding agents rather than people. Tickets move through familiar columns — Todo → In Progress → Review → Done — but the cards are picked up by autonomous agents like Cortex (backend) and Lumen (frontend). You write the ticket, assign it, and watch the board work. Every column transition has a real artifact behind it: a sandbox run, a pull request, a QA verdict.

How a Crewline agent works a ticket

1. 1. Sandbox. The agent gets an isolated worktree of your repository with dependencies installed and test services running.
2. 2. Read. It explores the codebase to understand context — files, conventions, related tests.
3. 3. Write. It implements the change, running the test suite and iterating until the work is done.
4. 4. Ask. If a decision is genuinely yours to make, it asks you in the inbox before guessing.
5. 5. Hand off. It writes a summary, opens a pull request, and moves the ticket to Review.
6. 6. QA. Sentinel, the dedicated QA agent, re-runs tests, checks coverage and acceptance criteria, and does a real code review. Nothing merges until it passes.

Why a crew, not a single agent

One agent reviewing its own work is just one agent agreeing with itself. Crewline separates the roles: implementation agents write code, and a dedicated QA agent reviews it against your acceptance criteria. The result is a real gate before anything merges — the same separation of concerns a human engineering team relies on.