For Anthropic CCA-F<\/a> candidates, multi-agent systems add separate context windows, separate tool loops, and coordination overhead. In Claude engineering<\/a>, they earn that complexity when independent workers can explore different parts of a problem in parallel, apply specialized prompts or tools, and return results that a lead process can synthesize.<\/p>\n Do not split a linear task just to say it is multi-agent. If every worker depends on the previous worker\u2019s full output, a sequential workflow or single agent usually costs less and is easier to debug.<\/p>\n Anthropic\u2019s own multi-agent research architecture uses an orchestrator-worker pattern for breadth-first research, where a lead agent delegates distinct searches to parallel subagents. The advantage comes from separation of concerns and independent context, not from the number of model calls.<\/p>\n Agentic orchestration<\/a> should choose single-agent, sequential, or parallel structure from the topology of the work rather than from a preference for multi-agent complexity. Parallel branches are valuable when the branches are genuinely separable.<\/p>\n A worker task should state the objective, scope, expected output, tools or sources allowed, and what not to duplicate. Vague delegation such as \u201cresearch this topic\u201d causes workers to overlap, chase different time horizons, or return incompatible levels of detail.<\/p>\n Give the lead agent a budget for workers, search\/tool calls, and time. Without a stopping rule, a coordinator can keep spawning subagents because each result exposes another possible branch.<\/p>\n Use the discipline from multi-step reasoning<\/a> to define completion criteria. The lead should know what evidence is enough to synthesize an answer and which unresolved gaps justify another worker.<\/p>\n Make delegation observable. Record which worker received which subtask and why, so duplicated effort and missed coverage can be diagnosed instead of blamed on vague model behavior.<\/p>\n Delegation should specify whether the worker is expected to discover evidence, transform a known artifact, challenge an assumption, or propose an action. These are different jobs with different stopping criteria. A worker assigned to verification should not silently broaden into open-ended brainstorming, and a worker assigned to ideation should not be treated as an authoritative verifier.<\/p>\n Use unique task identifiers and preserve parent-child relationships in the run record. When a coordinator revises or cancels a subtask, that state should be explicit rather than inferred from the latest prompt. This becomes important when workers complete out of order or when a retry produces a second candidate result for the same assignment.<\/p>\n Independent workers avoid filling one context window with every intermediate search or tool result. That allows each agent to focus on its assigned problem and can improve breadth on tasks where many independent directions matter.<\/p>\n The tradeoff is information loss at handoff. A worker that compresses ten sources into three bullets may omit caveats the lead later needs. Define output formats that preserve evidence references, confidence, unresolved conflicts, and any assumptions that affected the result.<\/p>\n Use durable artifacts for large outputs instead of copying everything through the coordinator. Files, structured records, or references can preserve fidelity while keeping the lead context compact.<\/p>\n Decide what shared memory is authoritative. Agents should not silently maintain incompatible versions of the plan or data model in separate contexts.<\/p>\n Design the worker return format around what the lead needs to decide. For research, that might include findings, source references, confidence, contradictions, and unresolved questions. For coding or data work, it may be better to return a file or patch plus a compact explanation rather than copying a large artifact through every agent context.<\/p>\n Shared facts need one authoritative store when they can change during the run. If every worker receives a snapshot and independently mutates its assumptions, the coordinator can end up synthesizing results based on different versions of the same input. Versioned artifacts and explicit read times make those conflicts visible.<\/p>\n One slow or failed worker should not block all useful results indefinitely. Track worker state separately and let the coordinator decide whether to retry, replace, continue with partial evidence, or stop.<\/p>\n Retries should preserve the original subtask and evidence of the failure. Spawning a new worker with a subtly different instruction can hide the fact that the system never solved the original problem.<\/p>\n Correlate worker spans through AI observability<\/a> so operators can see fan-out, latency, tool usage, result size, and failure propagation. A single top-level duration metric cannot explain why one branch consumed most of the budget.<\/p>\n Set limits on recursive delegation. A worker that can spawn more workers can create an uncontrolled tree unless depth, count, and cost are bounded.<\/p>\n Cancellation is a first-class case, not just an error. Once the lead has enough evidence, continuing every outstanding branch wastes latency and tokens; conversely, cancelling a worker that already triggered an external side effect can be unsafe. Give analysis-only workers and action-capable workers different cancellation policies.<\/p>\n Partial success should be represented structurally. A worker can return useful findings together with missing subquestions instead of forcing the coordinator to choose between ‘success’ and ‘failure.’ That lets the lead decide whether the gap is material to the user’s objective before spawning more work.<\/p>\n The lead agent should not concatenate worker outputs. It must reconcile overlaps, disagreements, evidence quality, and gaps, then produce a coherent answer that reflects the user\u2019s actual objective.<\/p>\n Require workers to cite or reference the evidence behind high-impact claims so the lead can compare sources instead of voting on unsupported summaries. When two workers disagree, preserve the conflict until a stronger source or evaluation resolves it.<\/p>\n Use structured synthesis inputs where possible: findings, evidence, confidence, risks, and open questions. This makes missing dimensions visible before the final response is written.<\/p>\n Evaluate synthesis separately from worker retrieval. A system can have excellent subagents and still fail because the lead drops important findings or overweights a confident but weak branch.<\/p>\n Multi-agent autonomy increases the number of opportunities for a mistaken action. Apply human oversight<\/a> where the system transitions from analysis to an irreversible external action, not necessarily on every internal delegation.<\/p>\n Workers can propose actions or evidence, while a central executor enforces authorization consistently. Do not allow specialized agents to bypass controls simply because they use a different prompt or toolset.<\/p>\n Make approval payloads include the provenance of the recommendation: which workers contributed, what evidence was used, and what unresolved uncertainty remains. A human cannot review a consequential synthesis if the path that produced it is hidden.<\/p>\n After denial, feed the reason into the coordinator so the system can revise the plan rather than repeatedly presenting the same action.<\/p>\n Every worker consumes model tokens, tool calls, and potentially external service cost. Multi-agent performance gains must be measured against this expanded resource footprint, especially for tasks that could be handled by one capable model call.<\/p>\nThe lead agent needs a delegation contract<\/h3>\n
Context separation is both a benefit and a source of loss<\/h3>\n
Parallel workers need independent failure handling<\/h3>\n
Synthesis is a distinct reasoning job<\/h3>\n
Human oversight belongs at high-consequence convergence points<\/h3>\n
Cost grows with breadth and duplicated context<\/h3>\n