One Agent, One Job
Give one agent one bounded outcome; add workers only for useful independent scopes.
Known intent and source owners define a bounded job. Factor IV sets the permissions and limits under which that job can run.
The rule
Give one agent one coherent outcome it can finish and check.
Keep one agent and one writer as the default. Several agents are useful when the work can be divided into independent scopes with a clear integration point.
Do not split a small task into roles merely because a factory can launch them. Keep continuity where it helps execution. Use a fresh context when independent judgment is required, rather than resetting context at every phase.
The evidencemeasured
Controlled comparisons of one agent against several test this rule directly.
- Kim et al., 2025 ran 260 configurations across six benchmarks. Against a single agent, multi-agent setups ranged from 80.8% better on financial reasoning that splits cleanly to 70.0% worse on sequential planning. On SWE-bench Verified, every multi-agent design scored below the single agent.
- CooperBench, 2026 assigned two coding agents separate features in one codebase, across more than 600 tasks. Their success rate working together was 30% lower than performing both tasks individually.
- Anthropic, 2025 reports that its multi-agent systems use about 15 times the tokens of a chat, and that most coding tasks have fewer truly parallel parts than research.
The counterexample is worth knowing. Cursor, 2026 reports running hundreds of agents on one codebase for weeks, after it separated planners from workers. Agents coordinating as equals had failed. The report gives scale and no acceptance rate.
Job size has a ceiling too. METR, 2025 measures the length of task a model completes half the time: about 50 minutes for the frontier models of early 2025, doubling roughly every seven months. Size the job to what the current agent can finish and check.
Put it to work
- Describe the output and acceptance criteria before choosing a runtime.
- Keep related source reads, drafting and ordinary checks together.
- Split only work with independent scopes and useful separate outputs.
- Assign integration when the caller has authorized multiple workers.
Example · the monthly operations report
One agent reads the ledger, drafts the monthly report and runs reconciliation. A separate reviewer is useful if the report contains consequential judgments that those checks cannot establish. Five agents rewriting the same draft add coordination without a separate outcome.
Failure signal
- Workers duplicate the same investigation or edit the same output.
- Coordination takes more effort than the task.
- Nobody owns the assembled result.
Done looks like
The assigned job has one accountable execution lane, a defined output and an appropriate check path. Any additional lane has a distinct purpose.