Orca: coordinate coding agents without losing track of their work
Measure useful Orca throughput, not the number of open agents
Compare a fixed task set while recording integration conflicts, model usage and review effort.
What you will learn
- Define completed work
- Include machine and model costs
- State what remains unmeasured
Before you start
- Basic command-line and configuration reading
- Ability to work in a disposable authorized environment
Create a learning tool that records host, worktree and acceptance evidence without automating merges or deleting unfinished work.
Key takeaways
- Accepted work is the useful output.
- Worktrees do not remove semantic conflicts.
- Model and human costs both matter.
Define completed work
Choose independent, labeled tasks with tests and acceptance conditions. Compare sequential execution with a small concurrent run using the same agent and model settings. Count only reviewed, accepted results as completed work.
Keep shared-interface changes out of the initial sample or label them separately. Separate worktrees do not prevent two agents from making incompatible assumptions about a schema or API. Integration failure belongs in the result.
Include machine and model costs
Record wall-clock completion, peak memory, CPU pressure and disk growth, plus usage from each agent account when available. Multiple worktrees may also duplicate dependencies or build outputs. Do not infer a saving from the number of terminals.
Measure time spent reading diffs, resolving conflicts and rerunning tests. Faster generation can still increase review effort. For remote trials, record network conditions and host resources instead of attributing every delay to the desktop.
State what remains unmeasured
Repeat enough trials to expose variability and report the task sample and failures. A throughput ratio from one repository should not be generalized to unrelated languages or task dependencies.
No performance experiment was executed here. The README’s ambitious productivity language is not our measurement. The proposed record leaves all observations empty until a controlled trial supplies them.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Label independent tasks and acceptance checks.
- 2
Compare sequential and small concurrent runs.
- 3
Count review, conflicts and resource usage.
Copy-ready example
{
"experimentProposal": true,
"tasks": null,
"concurrency": null,
"accepted": null,
"reviewMinutes": null,
"modelUsage": null,
"peakMemory": null,
"executed": false
}Frequently asked questions
Are more agents always faster?
Concurrency can add contention and integration work.
What should a speed claim include?
Task sample, settings, accepted outputs, failures and review effort.
Sources
- Orca / README.mdSource checked 2026-09-18
- Orca / package.jsonSource checked 2026-09-18