Features
Orchestration & Arena
Decompose a goal into a subtask plan, or race the same goal across several agents and judge the winner.
A single session is one agent working on one thing. Sometimes you want more: a big goal broken into a plan the agent works through in order, or several agents racing the same goal so you can pick the best result. Orchestration and Arena are the two multi-session workflows Operon gives you, and they solve different problems.
Orchestration
Orchestration takes one goal and decomposes it into a plan of subtasks, then works through them. You give Operon a goal; it produces a subtask breakdown with a recommended tool per subtask, which you can edit before starting. Once started, the plan drives sessions for each subtask and advances as they finish.
Note
Orchestration runs subtasks sequentially — one at a time, in order. Parallel execution with dependency-graph resolution is planned but not yet available, so today a plan is a reliable to-do list Operon works down, not a fan-out of concurrent agents.
- 1
Describe the goal
Enter a goal in the plan creator. Operon decomposes it into subtasks using keyword-based heuristics (add, fix, refactor, migrate, test, document) and recommends a tool for each.
- 2
Review and edit the plan
The plan is a draft you own — reorder, reword, or drop subtasks before you start. Nothing runs until you say so.
- 3
Start the plan
Operon works through the subtasks in order, launching a session for each. You can pause, resume, or cancel the whole plan at any point.
- 4
Watch it advance
As each subtask’s session ends, the plan advances to the next. Progress bars and expandable subtask lists show where things stand.
Arena
Arena is agent-versus-agent. You give the same goal to two, three, or four agents at once, each running in its own isolated worktree so they never step on each other. They race. Then Operon judges them on data nobody else has — cost, files changed, diff quality, scope violations — and lays out a side-by-side verdict strip so you can pick the best result.
- Same goal, different agents — spawn 2 to 4 agents on the identical prompt.
- Isolated worktrees — every seat gets its own git worktree, so their changes never collide.
- Live verdict strip — cost, files changed, diff-quality score, and status per seat, updating as they run.
- Judge on real data — the comparison is grounded in what each agent actually did, not vibes.
The advisory auto-judge
Once every seat has delivered, Operon can run an advisory judge that reads the structured signals — cost, diff quality, files changed, tool success — and recommends a winner with a short rationale and a confidence level. This is a recommendation only. It is deliberately built on structured metrics and short response excerpts, never on raw diffs, and it never touches the actual winner.
Important
The human always picks the Arena winner. The auto-judge is advisory — it recommends, you decide. When you pick a winner, that seat’s session continues into the normal review-and-ship flow, and the losing seats’ worktrees are reclaimed. Any uncommitted work in a loser’s worktree is preserved, not destroyed, so a good idea from a losing agent is never silently lost.
When to use which
Orchestration and Arena are not competing features — they answer different questions. Orchestration is for a big goal that has a natural order of steps. Arena is for a single goal where you are unsure which agent or approach will do best. And a plain session is still the right default for focused, ordinary work.
| Single session | Orchestration | Arena | |
|---|---|---|---|
| Shape | One agent, one goal | One goal split into subtasks | One goal, several agents |
| Concurrency | One session | Sequential subtasks | 2–4 concurrent seats |
| Isolation | Your working repo or a worktree | One session at a time | A worktree per seat |
| Outcome | The work itself | A completed plan | A judged head-to-head, human picks the winner |
| Best for | Focused, ordinary work | Multi-step goals with a natural order | Trying multiple agents on the same task |
Tip
Arena results are also viewable in the web app as a read-only verdict strip you can share with a teammate — the winner pick stays a desktop, human-only decision.