Overview
Orca coordinates parallel AI-assisted coding by binding each agent to a real Git worktree, an observable session, and a reviewable delivery path. The series moves from file isolation and status evidence to deliberate comparison, human review, structured Runs and Dispatches, and execution hosts for local, SSH, remote-server, or disposable environments. Its value is clearer ownership and safer recovery when alternatives, long-lived work, or remote execution make open terminals ambiguous. It does not make agents correct, worktrees sandboxed, remote silence proof of exit, or agreement a substitute for tests and human acceptance.
Learning outcomes
- Orca makes parallel attempts comparable by giving each agent a real worktree and explicit starting reference.
- Session status is evidence about process state, not proof that code is correct or ready to ship.
- Diff review, tests, hooks, and human acceptance remain necessary after an agent reports done.
- Remote execution and structured orchestration add reach and accountability, but also costs, permissions, and recovery boundaries.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What does Orca add to AI coding?
Orca adds a coordination layer connecting agents to Git worktrees, observable sessions, review, and delivery; it does not provide a smarter model.
When should developers run multiple AI coding agents in Orca?
Use multiple agents when competing approaches can resolve meaningful uncertainty and justify the added compute, setup, disk, and review effort; routine clear tasks may not benefit.
Does Orca’s worktree isolate an AI agent completely?
No. A worktree separates tracked files and branch history, but shared caches, credentials, network services, ignored state, and process permissions can still couple attempts.
What does a remote disconnect mean in Orca?
A disconnect makes the process unverifiable, not necessarily exited; the execution host must provide positive evidence before restart or cleanup.