Rethinking Agent Autonomy: A Two-Axis Model Beyond the Single Ladder

Software engineering is shifting from prompting AI agents to operating them, with software factories, subagents, hooks, and parallel agent fleets becoming everyday reality. Steve Yegge's widely-cited single-axis autonomy ladder gave engineers one number to gauge trust in an individual agent, but it can't capture the skill of coordinating many agents simultaneously — a capability that now carries as much weight as trusting one agent alone.

The fix is to separate two questions that autonomy debates usually conflate: how far can a single agent roam from human oversight (agency), and how well can you coordinate multiple agents at once (orchestration). Agency ranges from an agent merely suggesting actions, to working a scoped task with constant evidence-based reporting, to pursuing an open-ended goal through experimentation. Orchestration ranges from one agent on one thread, to several agents working in isolated worktrees, to a full orchestrator turning a backlog into continuous work — management by exception.

For engineers, the practical takeaway is choosing the autonomy level a task actually deserves and pairing it with defensible verification — passing tests, type checks, screenshots, or reproducible steps. Features like Claude Code's /plan, /goal, /loop, and /background modes, and Codex's Goal mode, Auto-review, and worktrees, already embed this two-axis thinking into shipping products. The biggest failure mode at scale is false parallelism — poorly decomposed tasks causing agents to collide rather than compound productivity.