When Multiple Agents Are Actually Justified

Agent count is an implementation choice, not a capability metric. A deterministic function is strongest when the path is known. A single model call can classify or transform bounded input. A fixed workflow can route, fan out, vote, or evaluate without giving a model control of the graph. One agent fits open-ended work inside one authority and state boundary. Multiple agents become a candidate only after the task exposes boundaries that one loop cannot handle cleanly.

These signals can combine, but none is automatic proof. Parallel calls may hit the same external bottleneck. Separate contexts may lose cross-branch facts. Credential isolation helps only when runtimes issue attenuated tokens and prevent lateral access. Organizational ownership can warrant a service boundary without warranting an autonomous agent on both sides.

• Require at least two work units that can begin without consuming one another's generated output.
• Make dependencies explicit in a DAG; a long serial chain is prompt chaining, even if every node is called an agent.
• Assign each mutable resource one writer or an explicit concurrency protocol. Natural-language consensus is not a lock.
• Let workers return bounded typed evidence; the aggregator owns deduplication, disagreement, missing branches, and final acceptance.

Context pressure is a capacity argument only when the task can be partitioned. Three workers each receiving the same giant dossier triple input work without expanding useful context. A better design gives each worker a bounded evidence shard and requires citations or content identities in its output so the merge can recover provenance rather than trusting a lossy narrative.

First-party engineering reports can motivate a hypothesis, not set your release bar. Anthropic describes a research product where parallel subagents explore independent directions and also reports materially higher token use; that product-specific observation does not establish that a multi-agent topology improves coding, support, or your task distribution. Reproduce the comparison locally with the same models, tools, data, timeouts, and stopping rules.

1def audit_decomposition(
2    contract: DecisionContract, record: DecompositionRecord
3) -> DecisionReport:
4    """Decide whether a bounded fork/join plan justifies separate agent contexts."""
5    contract = validate_record(contract, DecisionContract)
6    record = validate_record(record, DecompositionRecord)
7    if record.scope != contract.scope or record.contract_content_id != contract.content_id:
8        raise ValueError("decomposition belongs to another decision contract")

example=illustrative_only
decision=MULTI_AGENT_JUSTIFIED
branches=3;critical_path_units=2
signals=parallel-frontier,context-pressure,credential-isolation,bounded-coordination
estimated_tokens=27000;coordination_ratio=0.111
claim=LOCAL_DECISION_RECORD_NOT_MULTI_AGENT_BENCHMARK

The invented supplier-risk fixture has three independent read-only branches for policy, finance, and security evidence. Each has a distinct owner, tool domain, credential isolation key, context estimate, and typed finding schema. A fourth read-only unit depends on all three and emits the decision record. Their declared branch context exceeds one local context budget while each branch fits; the merge is below the contract's coordination-ratio threshold.

The audit accepts only this bounded fork/join shape. It reconstructs frozen records, checks exact contract scope, unique contiguous positions, backward dependencies, shared mutable isolation, a complete read-only merge, worker and context limits, finite exact-float ratios, exact booleans and strings, constructor bypass, and collection copying. Branches may share an isolation key only when every branch sharing it is read-only; mutable collisions require an explicitly serialized design outside this flat fork. `MULTI_AGENT_JUSTIFIED` is the result of an invented heuristic over estimates. The artifact runs no agent, measures no latency or quality, prices no tokens, and makes no benchmark claim.

1. 01Contract the branchDeclare objective, admissible evidence, context cap, tools, credentials, side effects, deadline, output schema, and completion test.
2. 02Bind the taskGive each worker a task and evidence digest so stale or cross-run output cannot enter the merge.
3. 03Collect partial outcomesDistinguish success, abstention, timeout, policy denial, and malformed output; do not let silence masquerade as no finding.
4. 04Resolve conflicts deterministicallyDefine evidence priority, duplicate identity, disagreement escalation, quorum if justified, and the sole owner of final acceptance.
5. 05Stop the graphCap worker count, depth, retries, tokens, tool calls, and elapsed time; cancellation must reach children and their external effects.

Retries need one owner. If both orchestrator and worker retry the same branch, multiplicative attempts can exceed budgets and repeat effects. A child that can write needs its own idempotency and recovery contract; a parent retry does not erase an ambiguous child outcome. Persist lineage from final claims to branch evidence and tool receipts so a bad synthesis can be separated from a bad source.

Use paired tasks and identical acceptance tests. Hold the model family, tool versions, corpus snapshot, budgets, and timeout policy constant enough to attribute the effect of topology. Report confidence intervals or repeated-run distributions when sampling variance matters. A win on one curated task is not a universal multi-agent advantage, and an internal vendor evaluation is not a transferable baseline.