Bounded Fan-Out

How many agents you spin up is a design variable with its own costs — a fixed, small roster and an unbounded, dynamically-sized swarm are two different operating points, and more agents is not monotonically better. A bounded fan-out fixes the roster up front: N named workers, known before the run, the same every time (STORM's five expert lenses + a fixed verification pass ≈ 12 agents). An unbounded fan-out lets an orchestrator decide the count at runtime and can balloon into hundreds (the source's Claude Deep Research run spawned "103 agents"). The bounded operating point buys predictable cost, predictable latency, a consistent output shape, and low provider-rate-limit exposure; the unbounded point buys breadth on topics whose sub-questions you can't enumerate in advance, at the price of cost/latency variance and rate-limit risk.

This is a distinct axis from Model-Tier Routing (which tier each agent runs, not how many) and from Agent Communication Topology (whether the agents can talk to each other). It sharpens the Agent Task Graph observation that "more agents, more autonomy, more compute do not always mean better outcomes" into a sizing decision: pick the roster cardinality deliberately, and default small when the task's angles are known.

Claims

Agent count is a first-class design variable, not a free "more is better" dial — a bounded fixed roster and an unbounded dynamic swarm are different operating points on cost, latency, rate-limit exposure, and output consistency. principle — durable: added parallelism carries cost and error surface, so the count is a choice to justify, the same discipline Agentic Simplicity and Agent Task Graph apply to team size.
Prefer a bounded fixed roster when the task's relevant angles can be enumerated up front and you want predictable cost and a consistent output shape. best practice — context: research and synthesis tasks with a stable decomposition (a fixed panel of expert lenses covers the topic). The bound is what makes the pipeline packageable and repeatable — "always your five personas." Where the needed sub-questions genuinely can't be predicted, an unbounded orchestrator-workers swarm that decides subtasks at runtime is the right tool (Agentic Workflow Patterns orchestrator-workers).
A large simultaneous fan-out risks provider rate limits that a small fixed roster sidesteps. best practice — context: API-metered agent runs; the source's 100+-agent Deep Research run "got hit by API rate limits," whereas a fixed five-persona roster is bounded by construction. The mitigation is real but contingent on the provider's concurrency limits, not a law.
Bounded need not mean lower quality: a small fixed roster with a verification pass can match or beat an unbounded swarm's output. observation — the source's single-run comparison had an external Codex judge rate the ~12-agent STORM briefing above the 100+-agent Deep Research report on six dimensions. A single anecdote, not a controlled benchmark; the "faster and 100% cheaper" figure is the source's claim, not a measurement. See LLM-as-Judge.
Agent count and attempt count are different dials: this page's "more is not monotonically better" governs the roster, not retries against a checker. principle — durable, and the boundary that keeps this page compatible with Repeated-Sampling Scaling. A second source (Nate B Jones) reports a study where the same model at 250 attempts beat the best single attempt of the best available model — smooth improvement across four orders of magnitude — while this page's cost, latency, and rate-limit objections to fan-out all still stand. The two claims coexist because they are about different things: parallel breadth across agents versus sequential persistence against a verifier. The failure this creates is reading a sampling result as licence to spin up a swarm; the source is explicit that a swarm without a verifier stalls where a single checked loop wouldn't. Figures are the source's claims — see that distillate's Key claims.
The prior question to cardinality is whether to form a team at all, and it has a short answer. best practice — context: any fan-out proposal. Before sizing the roster, name which of the two limits — capacity or separation of concerns — the team answers; a roster answering neither is "just more agents" (Team-Forming Constraints). This page's "default small when the task's angles are known" is the sizing rule that follows once the team is justified.

Adversarial Planning Council — a council of a fixed number of opposed personas is a bounded fan-out; the bound is what lets it ship as a drop-in, always-the-same-roster file.
Agent Task Graph — "more agents ≠ better; what matters is coordination toward a common goal" is the coordination-side statement of this sizing discipline; bounded fan-out is the cardinality choice.
Agentic Workflow Patterns — the axis maps onto two named patterns: parallelization (fixed, pre-defined subtasks = bounded) vs orchestrator-workers (subtasks decided at runtime = unbounded).
Model-Tier Routing — the orthogonal fan-out lever: which tier each agent carries, decided independently of how many agents there are. Running fewer agents and running them cheaper are different economies.
Agent Communication Topology — the third fan-out axis: whether the roster is hub-and-spoke or a mesh. Count, tier, and topology are chosen separately.
Workflows vs Agents — a fixed roster is workflow-shaped (predefined structure); an unbounded runtime-sized swarm is agent-shaped (the model directs the decomposition).
Recency-Grounded Research — a sibling bounded pattern: sub-agents fanned across a fixed set of named sources inside a bounded recency window.
Agentic Simplicity — "default to the smallest roster that covers the task" is the simplicity discipline applied to fan-out cardinality.
Team-Forming Constraints — the prior, qualitative question: whether a team is justified and by which of two limits. This page answers how many once it is; the two compose as gate then sizing.
Agent-Shape Triage — one step earlier still: the pre-flight estimate that decides whether the task wants chat, one agent, a team, or nothing at all.
Repeated-Sampling Scaling — the dial that does scale, and the reason this page's claim must be read as being about roster size rather than about persistence.
Distillate: STORM: A Fixed Panel of Adversarial Research Lenses, Packaged as a Skill
Distillate: 1.6M agents registered for OpenClaw and did NOTHING. — draws the attempts-vs-agents boundary and supplies the prior gate (name the limit your team answers) that sits above cardinality.

Bounded Fan-Out

Claims

Related

Linked from