6.b.iii Use multi-agent orchestration only for breadth-first, independent sub-tasks

Spawn parallel sub-agents when the work splits into independent, breadth-first parts with little shared context; expect about 15x the tokens of a chat and up to 90% less wall-clock. Default to a single agent otherwise (most coding): parallel workers make conflicting implicit decisions, and decompose-first planning scored below a single agent on data questions. When delegating, share full traces, give each sub-agent an objective, output format, tools and boundaries, have it write outputs to storage and return references, and re-verify what it reports.

Tier 4 because a multi-agent topology is a restructuring; only a production test shows its behavior on live traffic.

When it applies

In the code: a multi-agent design whose sub-tasks share context or depend on each other; or a single agent on a breadth-first task with many independent parts
In traces: sub-agents duplicating work or making conflicting decisions; failure rows labeled delegation-failure-subagent or communication-failure-subagent; token use per task far above a single agent's
In eval results: single-agent against multi-agent on the same tasks

How

Check the task shape: independent parts, little shared context.
If so, prototype a lead plus 3-5 parallel sub-agents with the delegation rules above.
Production A/B against the single agent on cost, latency and quality.

How to measure it

production test (tier 4): A/B, single agent against multi-agent

To compare the change against your agent's current version, with the same cases and enough repeats to see past run-to-run noise, use aot-evals.

Expected impact

Quality: mixed, large: +90.2% on breadth-first research (Jun 2025); below a single agent when sub-tasks depend on each other (GATE, Jun 2026)
Cost: degrades, large: about 15x the tokens of chat
Performance: improves, large: up to 90% less time with parallelism

Evidence

Evidence comes from: Web research, Code generation, Data analysis.

Orchestrator-worker with capable lead, cheaper parallel subagents

"a multi-agent system with Claude Opus 4 as the lead agent and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval" (n not given; "excel especially for breadth-first queries"). Cost: "agents typically use about 4x more tokens than chat interactions, and multi-agent systems use about 15x more tokens than chats". Poor fit: "domains that require all agents to share the same context or involve many dependencies between agents ... most coding tasks".

Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic), "How we built our multi-agent research system" · Jun 2025Older evidence · Benefits of a multi-agent system; Architecture overview for Research
Specify each subagent task with objective, format, tools, boundariesNot measured

Short instructions like "research the semiconductor shortage" led subagents to misinterpret or duplicate ("one subagent explored the 2021 automotive chip crisis while 2 others duplicated work investigating current 2025 supply chains"). "Without detailed task descriptions, agents duplicate work, leave gaps, or fail to find necessary information."

Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic), "How we built our multi-agent research system" · Jun 2025Older evidence · Prompt engineering: Teach the orchestrator how to delegate
Spawn subagents in parallel, not serially

Together with parallel tool calls in subagents: "These changes cut research time by up to 90% for complex queries". Joint figure for both changes.

Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic), "How we built our multi-agent research system" · Jun 2025Older evidence · Prompt engineering: Parallel tool calling transforms speed and performance
Issue 3+ tool calls in parallel per subagent

Joint figure with parallel subagent spawning: "cut research time by up to 90% for complex queries".

Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic), "How we built our multi-agent research system" · Jun 2025Older evidence · Prompt engineering: Parallel tool calling transforms speed and performance
Run subagents asynchronously (proposed)Not measured

Currently synchronous: "the lead agent can't steer subagents, subagents can't coordinate, and the entire system can be blocked while waiting for a single subagent". Async "adds challenges in result coordination, state consistency, and error propagation".

Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic), "How we built our multi-agent research system" · Jun 2025Older evidence · Production reliability: Synchronous execution creates bottlenecks
Hand off to a fresh subagent near the context limitNot measured

"When context limits approach, agents can spawn fresh subagents with clean contexts while maintaining continuity through careful handoffs."

Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic), "How we built our multi-agent research system" · Jun 2025Older evidence · Appendix: Long-horizon conversation management
Have subagents write outputs to storage, return referencesNot measured

"This prevents information loss during multi-stage processing and reduces token overhead from copying large outputs through conversation history." Best "for structured outputs like code, reports, or data visualizations".

Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic), "How we built our multi-agent research system" · Jun 2025Older evidence · Appendix: Subagent output to a filesystem
Default to a single-threaded linear agentNot measured

"The simplest way to follow the principles is to just use a single-threaded linear agent ... Here, the context is continuous." "the simple architecture will get you very far".

Walden Yan (Cognition), "Don't Build Multi-Agents" · Jun 2025Older evidence · A Theory of Building Long-running Agents (after Principle 2)
Share full agent traces with subagents, not task messagesNot measured

Principle 1: "Share context, and share full agent traces, not just individual messages." Copying the original task is not enough because "the conversation is most likely multi-turn, the agent probably had to make some tool calls to decide how to break down the task". Illustrated with a Flappy Bird clone whose subtasks were each misread.

Walden Yan (Cognition), "Don't Build Multi-Agents" · Jun 2025Older evidence · Principle 1
Do not run parallel agents that make implicit decisionsNot measured

Principle 2: "Actions carry implicit decisions, and conflicting decisions carry bad results." Even with shared context, parallel subagents produced "a bird and background with completely different visual styles". Principles 1 and 2 are "so rarely worth violating, that you should by default rule out any agent architectures that don't abide by them."

Walden Yan (Cognition), "Don't Build Multi-Agents" · Jun 2025Older evidence · Principle 2
Limit subagents to answering well-defined questions, sequentiallyNot measured

Claude Code (June 2025) "never does work in parallel with the subtask agent, and the subtask agent is usually only tasked with answering a question, not writing any code ... The benefit ... is that all the subagent's investigative work does not need to remain in the history of the main agent".

Walden Yan (Cognition), "Don't Build Multi-Agents" · Jun 2025Older evidence · Applying the Principles: Claude Code Subagents
Decide and apply edits in one model actionNot measured

Edit-apply systems "would still be very faulty ... the small model would misinterpret the instructions of the large model and make an incorrect edit due to the most slight ambiguities". "Today, the edit decision-making and applying are more often done by a single model in one action."

Walden Yan (Cognition), "Don't Build Multi-Agents" · Jun 2025Older evidence · Applying the Principles: Edit Apply Models
Avoid agents negotiating with each other to resolve conflictsNot measured

"running multiple agents in collaboration only results in fragile systems. The decision-making ends up being too dispersed and context isn't able to be shared thoroughly enough between the agents."

Walden Yan (Cognition), "Don't Build Multi-Agents" · Jun 2025Older evidence · Applying the Principles: Multi-Agents
Require subagent delegation for large review corpora

Controlled experiment, 6 models x 5 scenarios x 20 runs per condition (1,200 runs), prohibited -> required delegation: mean files touched 86.9% -> 97.3%, lines read 67.0% -> 87.3%, needles reported 49.9% -> 69.6% (e.g. Sonnet 5 27.7% -> 53.7%, Opus 5 69.7% -> 95.7%). Explicit overclaiming across all runs 34.5% -> 16.3%. All-files-touched increase significant in both families (Claude G2=46.27, GPT G2=39.58, p<0.0001).

Smyth, Mantilla-Ramos, Tikeng Notsawo, Helbling, Tosato, Merzouk, Dziri, Gidel, Tosato (Tara Research, Mila, Cohere), "Quantifying Overclaiming Propensity in Frontier LLM Agents" · Sep 2026 · §3.4 Controlled delegation experiment, §4, Appendix D.2 Tables 4-5, Figure 4
Delegation does not make coverage reports honest

Same 1,200-run experiment: among incomplete reviews, misleading rate rose 80.7% -> 93.8% overall (Claude family condition effect G2=19.10, p<0.0001; Fable 5 64.9% -> 100%, Sonnet 5 60.8% -> 84.2%); GPT models stayed at or near 100% in both conditions. 50.3% of incomplete delegating reviews still explicitly overclaimed; 83-100% failed to disclose the gap across the six models.

Smyth, Mantilla-Ramos, Tikeng Notsawo, Helbling, Tosato, Merzouk, Dziri, Gidel, Tosato (Tara Research, Mila, Cohere), "Quantifying Overclaiming Propensity in Frontier LLM Agents" · Sep 2026 · §4 Subagents increase coverage but do not improve honesty, Appendix D.2 Table 5
Decompose the query into sub-tasks before observing data

Table 1, decomposition-first ReDel vs single-agent ReAct: Qwen3-8B EHRSQL 26.4 vs 30.2, LS-Hard 8.0 vs 14.5; GPT-5.4-mini EHRSQL 26.8 vs 32.2. ReDel beats ReAct only on RealEHR (both models) and LS-Hard GPT-5.4-mini (40.2 vs 37.8). Authors argue decomposition is fixed "before sufficient database observations are available."

Youngwon Lee, Jaejin Kim, Seung-won Hwang (Seoul National University), "Bootstrapping Semantic Layer from Execution for Text-to-SQL (GATE: Grounding After Test from Execution)" · Jun 2026 · §2 Decomposition for grounding; Table 1
Stay single-agent until breadth-first exploration pays offNot measured

Rec 12 citing [17]: multi-agent systems use "~15x more tokens" than a chat baseline and "most coding tasks involve fewer truly parallelizable tasks than research".

Barbaste et al., "Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents -- A Source-Code Study of Eleven Systems" · Jul 2026 · Recommendation 12; Sec 15.1
Spawn sub-agents with isolated context, summary-only returnsNot measured

Claude Code AgentTool agent definitions (Explore, Plan, VerificationAgent); OpenCode returns "only the final assistant text wrapped in <task_result>"; Hermes "summary-only return"; Pi separate OS process "hard context isolation by construction".

Barbaste et al., "Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents -- A Source-Code Study of Eleven Systems" · Jul 2026 · Sec 11.1–11.9; Table 11 "Recursive Composition"
Control how much history a sub-agent inheritsNot measured

Codex SpawnAgentForkMode "(FullHistory or LastNTurns(N))"; keep_forked_rollout_item() "preserves only system/developer/user messages and final assistant answers, filtering out intermediate tool calls and reasoning."

Barbaste et al., "Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents -- A Source-Code Study of Eleven Systems" · Jul 2026 · Sec 11.3; Sec 6.4
Restrict sub-agent tools and block recursionNot measured

Claude Code workers restricted to a 16-tool whitelist; Hermes "subagent must not gain tools the parent lacks", six-tool blocklist, defaults "three concurrent children, 50-iteration budgets, depth 1"; Mistral Vibe only "subagent"-tagged profiles (explore) are delegable; OpenCode children inherit the parent's deny rules and recursion is off by default.

Barbaste et al., "Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents -- A Source-Code Study of Eleven Systems" · Jul 2026 · Sec 11.2.3; Sec 11.5; Sec 11.7; Sec 11.9
Quarantine subtasks in separate sub-agent contexts

Quotes Anthropic's multi-agent research system: Opus 4 lead with Sonnet 4 subagents "outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval". Secondary citation.

Drew Breunig, "How Long Contexts Fail / How to Fix Your Context", https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html ; https://www.dbreunig.com/2025/06/26/how-to-fix-your-context.html (2025-06-26) · Jun 2025Older evidence · 06-26 post, "Context Quarantine"
Fan out independent checks to isolated sub-agentsNot measured

Claude Code team "uses sub-agents in code review to independently check different possible issues. A similar map-reduce pattern can be used for tasks like updating lint rules or migrations."

Lance Martin, "Agent Design Patterns" · Jan 2026 · Isolate Context

Also known as

Parallelization by sectioning or voting
Orchestrator-workers

5.c.iiSub-agent architectures with clean contexts
6.b.iiLLM map-reduce
9.b.iiiVerify completion against the environment before the agent stops
6.c.iCode-then-execute (CaMeL)
7.d.iContinuous autonomous task loop
4.b.iiInitializer-maintainer dual agent
1.b.iMatch the model and its effort to each call site and query class
7.c.viMulti-agent debate
6.a.iOracle-worker multi-model
6.b.iParallel tool execution for read-only calls

In Claude Code

This optimization is a skill in aot-kit, our open-source Claude Code plugin: /aot-kit:multi-agent-for-breadth. Its library entry holds every source. Have new evidence, or a correction? Suggest a change.

Use multi-agent orchestration only for breadth-first, independent sub-tasks