Justify the cost of splitting work

Learning objective

Decide when multiple agents outperform one by weighing parallelism and context isolation against token cost and coordination overhead.

Core idea

For an architect funding a production system, multiple agents earn their place when the additional accepted work justifies the additional inference and coordination. Specialist selection is assumed here. The architectural question is whether separate execution contexts improve the deliverable enough to meet a business constraint that one agent misses.

Anthropic identifies independent breadth, information beyond one context window, and complex tool sets as useful multi-agent workloads. Its research account also warns about tightly coupled work and shared context. Compare architectures on the same acceptance criteria: a faster incomplete answer is not an improvement, and more tokens are not evidence of better coverage.

How it works

Identify the benefit you expect

Parallelism helps when workers can begin with complete inputs and investigate independent branches. The Agent SDK documents concurrent subagents finishing independent work in the time of the slowest branch rather than the sum. That describes the parallel investigation phase. Planning, dispatch, result checks, synthesis, and recovery still contribute to end-to-end latency.

Separate elapsed time from total computation. Four concurrent workers can finish sooner while consuming more aggregate inference. If each must repeatedly ask the coordinator for another worker’s findings, the apparent fan-out contains dependencies. Measure waiting and repeated investigation before treating that shape as a latency improvement. See 1.4.2 for topology choices.

Context isolation helps when a branch reads substantially more than the next stage needs. SDK subagents keep intermediate tool calls and results in their own conversations and return a final message. This permits a worker to inspect long records while the coordinator receives an evidence-bearing summary. A fork inherits the parent conversation, so choose fresh-context investigation when filtering branch-specific material is the intended benefit.

Isolation also creates a handoff cost. The coordinator must receive enough evidence to verify the result. If the summary discards the deciding qualification, reopening sources or commissioning corrections can erase the saving. Measure how often results require reconstruction. The contract in 1.4.3 makes the useful compression explicit.

Account for all the inference

Anthropic’s research account reports about 4 times the tokens of chat interactions for agents and about 15 times for multi-agent systems. Both comparisons use chats as their baseline. They are observations from that system, not a promised multiplier for replacing your single agent. The same account reports that token usage alone explained 80% of performance variance in its BrowseComp analysis, supporting the economic case for measuring additional computation alongside quality rather than crediting agent count alone.

Budget the lead’s planning and synthesis, every worker’s requests, and recovery. In the Agent SDK, subagent requests count toward the query’s total_cost_usd. Compare cost per accepted deliverable alongside completion time and supported coverage. Counting only worker output tokens misses input processing and the coordinator’s additional work; counting only successful runs hides the cost of rejected reports.

Use representative cases where one agent is already competently prompted and equipped. Include the supported on-demand tool-search strategy in 3.1.3 for a large catalog and the compaction or tool-result-clearing strategy in 2.4.3 for growing context when they address the observed bottleneck. Anthropic’s guidance starts with the simplest solution and adds complexity when outcomes demonstrate its value. A weak single-agent baseline makes unnecessary orchestration look economical. Compare equivalent source access, acceptance rules, and task scope before attributing a gain to agent count.

Keep the decision narrow. If the benefit appears only for broad investigations, route those investigations through multiple agents while retaining one-agent execution for focused requests. Mandatory acceptance checks still apply to both routes; see 1.1.2. The cost decision should not silently remove a required review.

Bound the experiment and the service

The SDK exposes maxBudgetUsd in TypeScript and max_budget_usd in Python. The cap covers subagent spend, refuses further spawning with Budget limit reached, stops running background subagents, and ends the query with error_max_budget_usd. Those enforcement behaviors require Claude Code v2.1.217 or later.

This is an estimated-spend stopping threshold, not an exact invoice ceiling. Claude Code checks cost after a response arrives. Spend can pass the threshold by that response’s cost plus expenditure from subagents before they stop. Leave headroom below the business limit and include capped runs among unsuccessful outcomes. See 1.4.4 for preserving useful work when execution stops.

The platform owner sets and verifies these controls; the product owner decides whether the accepted-result improvement is worth funding. Record the evaluated task class, evidence-quality gate, latency target, observed spend, and conditions for returning to the simpler route. This decision record makes a request to add agents reviewable rather than an open-ended promise of intelligence.

Key distinctions and trade-offs

The expected benefit must survive its associated cost:

Token efficiency and business efficiency are different. A higher-token run can replace valuable manual work, but that conclusion needs accepted output and a defensible resource envelope. Parallelism can improve latency without lowering spend; isolation can improve accuracy without shortening the run. State which benefit pays for the architecture.

Architecture / decision scenario

A retailer needs a supplier-disruption brief covering twelve independent product categories before a morning procurement meeting. The business requires attributable findings within ten minutes and sets a hypothetical operating allowance of twenty dollars per accepted brief. No brief can omit a category silently.

The one-agent baseline meets accuracy checks but misses the ten-minute target because category research runs sequentially. You test concurrent category investigations with fresh contexts and evidence-bearing final reports. Cross-category purchasing recommendations remain a dependent synthesis stage. You measure total elapsed time, accepted category coverage, correction requests, and total_cost_usd across successful and failed runs.

The split earns deployment only if complete, supported briefs meet the target within the allowance after including synthesis and recovery. Set the SDK spend threshold below twenty dollars to leave documented overrun headroom. A capped run preserves its verified category findings but does not count as an accepted complete brief. If you resume it later, fund the resumed call from the remaining business allowance and account for cumulative session spend as described in 1.4.4.

A separate request asks about one supplier’s documented lead time. It stays on the one-agent route because the parallel breadth and context-filtering advantages do not apply. If broad runs start duplicating searches or requiring extensive source reconstruction, the architecture owner revisits the measured benefit rather than increasing agent count by default.

Common exam traps

• Treating the research account’s 15-times-chat observation as a universal single-agent multiplier.
• Measuring parallel worker time while excluding planning, synthesis, and recovery.
• Assuming separate conversations automatically produce adequate evidence handoffs.
• Comparing a tuned multi-agent system against an unnecessarily weak single agent.
• Treating maxBudgetUsd as an exact billing ceiling.
• Paying for the broad-investigation route on every focused request.

What to remember

Split work when independent execution or context filtering improves accepted outcomes under the actual business constraints. Measure the entire run, including rejected output and recovery. Defend the additional spend with a demonstrated benefit, and retain the simpler route where that benefit disappears.

Knowledge check

References

• Anthropic, How we built our multi-agent research system
• Anthropic, Building effective agents
• Anthropic, Subagents in the SDK
• Anthropic, How the agent loop works
• Anthropic, Tool search tool
• Anthropic, Context editing
• Anthropic, Track cost and usage