# Single agent versus parallel agents — October 21, 2026 (ET)

This is an owned, synthetic code-synthesis experiment, not a benchmark of production coding throughput. Real GPT-6.1-sol agents used medium reasoning effort with fresh context. No customer data, repository changes or production access was given to the experiment agents.

Download `swarm-experiment-2026-10-21.json` and `swarm-experiment-replay.mjs` from the same downloads directory. In Node.js, run:

```sh
node swarm-experiment-replay.mjs swarm-experiment-2026-10-21.json
```

This replays acceptance against the published owned sources; it does not invoke a model, spend inference credits, or reproduce the original elapsed times. Node `vm` is an execution convenience, not a security boundary: replay only this inspected owned corpus, not arbitrary downloaded source.

## Fixed protocol

The source and case vectors were written before dispatch. The JSON contains the task requirements, exact per-agent dispatch prompt and prompt-file payload, generated source, host timestamps, completion reply and case-level outcomes. The prompt-file paths are temporary fixture names; they contain no credentials. Each worker could read its own prompt and write its own submission only. Agents were told not to inspect tests or other submissions and not to delegate.

The three tasks are: normalize string tags without mutating input; paginate an array with validated positive safe integers; and remove sensitive header names using case-insensitive matching. Requirements and all 18 case vectors appear in the JSON. No libraries, network access or side effects are part of the functions.

Block 1 ran one agent for all tasks, then three agents in parallel with one task each. Block 2 reversed that condition order. Every agent used fresh context and the same model/effort setting. The coordinator dispatched parallel agents in separate tool calls, so dispatch was staggered rather than perfectly simultaneous. One accepted task means every assigned case passed, including the check that input was not mutated. Each case has a 50 ms execution timeout. Acceptance code does not retry a failed case. No repair redispatch was permitted; a failure would remain in the result.

## Accounting

- **Elapsed time:** host timestamp before dispatch through the latest submission file modification time. Includes dispatch/startup, prompt reading, synthesis, and file write. Excludes fixture preparation, acceptance replay and review. It is not provider inference latency.
- **Repair retries:** new agent dispatches to repair an observed acceptance failure. None were allowed or performed. Internal model/tool retries were not instrumented and are unknown.
- **Billed cost:** `null`. No per-call invoice or attributable usage receipt was available. Agent-call counts do not establish dollar cost.
- **Tokens:** unmeasured; no token total or list-price estimate is published.
- **Human review minutes:** `null`. No timed human review occurred. Machine checks do not measure human effort.
- **Coordination overhead:** dispatch and submission overhead stays in elapsed time. Harness creation, article writing and acceptance replay are excluded. These are not measurements of total project delivery effort.

## Results and failures

All 12 submitted task outcomes passed, across four conditions and eight agent invocations. Each condition ran the same 18 cases. The exact source for every outcome is preserved. There were no failed submissions, and no repair loop. Harness selftests separately reject a deliberately wrong function, missing source and nonterminating code; these are harness controls, not failed agent outcomes.

Single-agent elapsed times were 19.753 and 18.901 seconds. Parallel elapsed times were 22.165 and 26.908 seconds, rounded to milliseconds. These observations do not establish that one strategy is generally faster, cheaper, safer or more reliable.

The sample is small, host load was not controlled, and provider queue/cache state was unknown. Tasks were independent, short and output-only. There was no merge conflict, shared interface, browser flow, deployment or failed-task recovery. Repeat on your own representative tasks; preserve failures, measure attributable billed usage and timed human review, and define acceptance before dispatch. Do not extend these timings into a projected production saving.
