Coordinate Multiple Agents Only When It Helps

Divide independent work, preserve its evidence, and measure the cost of putting it back together.

A learning center wants to compare three public resources for teaching basic statistics. One agent could inspect them in sequence. Three agents could inspect one resource each. Ten agents could produce a much larger pile of notes.

Which arrangement is better?

The answer depends on the work, the quality of the outputs, and the effort required to reconcile them. More workers can reduce elapsed time when tasks are independent. They can also duplicate searches, repeat mistakes, and leave the final editor with more work than before.

Find the work that is actually independent

A useful split has clear boundaries and outputs that can be checked separately. For the resource comparison, each worker can inspect one assigned resource for coverage, prerequisites, worked examples, and source references.

The final recommendation depends on all three reviews, so it belongs to a synthesis step. Assigning every worker to “find the best statistics resource” would create overlapping searches and inconsistent comparisons.

Tasks with many shared decisions are harder to parallelize. If one worker is changing a data schema while another builds queries against it, both need a settled interface. If every worker needs the entire evolving context, coordination can consume the expected gain.

Anthropic's account of its multi-agent research system describes benefits for parallel research along with increased token use and coordination demands. Those observations concern its system and tasks; they are not a universal performance multiplier for every agent team.

Write an assignment contract for each worker

An assignment should name the scope, allowed sources, required evidence, budget, and stopping condition. The worker should know what belongs elsewhere.

For a controlled classroom comparison, use a fixed packet of public-source excerpts with known permissions and retrieval dates. Both the single-worker and multiple-worker conditions receive the same packet. This avoids giving one condition a larger or newer evidence base.

A worker record might contain:

The synthesis owner receives comparable records rather than three unrelated essays.

Give shared artifacts one owner

Workers can write separate result files or records. The synthesis owner merges accepted results into the final comparison. Avoid multiple agents rewriting the same final document simultaneously unless the system has a tested coordination mechanism.

Shared state should distinguish assigned, running, complete, failed, and canceled work. A worker's partial note can be useful without being marked complete. An expired assignment should not continue consuming resources unnoticed.

Keep the original evidence references with each claim. A synthesis agent should not have to trust another agent's confident summary when the supporting passage is available. If the passage is missing, the claim remains unverified.

A model-generated statement that “the other worker approved this” is not an approval record. Preserve the same authority boundaries used for single-agent work.

Compare quality and cost on the same tasks

Run the single-worker baseline and the proposed team on equivalent task sets. Include synthesis, review, retries, and failed workers in the measurements.

Use several representative tasks rather than one demonstration. A broad literature comparison may benefit from parallel reading, while a tightly dependent debugging task may not. Report results by task type so an overall average does not hide the difference.

The following numbers are fictional planning data, not results of a live agent experiment:

Assuming the three workers truly run concurrently and review begins afterward, the team saves five minutes of elapsed time. It also uses 15 worker-minutes rather than 12, more review time, and a higher model cost. The acceptance count is unchanged.

Whether the trade-off is worthwhile depends on the value of the earlier result. Do not call it an efficiency improvement without saying which resource improved and which became more expensive.

Keep parallel speedup realistic

If part of a job must remain sequential, adding workers cannot eliminate that part. Suppose 40% of a task is inherently sequential and 60% can be divided perfectly among three workers. The ideal normalized time is 0.4 + 0.6/3 = 0.6, giving an ideal speedup of about 1.67 times before coordination overhead.

That is a mathematical illustration, not a prediction for language-model agents. Real tasks add variable response times, rate limits, repeated work, and synthesis costs. The slowest required worker can determine when the final stage begins.

Use this calculation to question extravagant claims, then measure the actual workflow.

Resolve disagreements through evidence

Suppose one worker reports that a resource requires algebra and another says no prerequisites are needed. First check whether they reviewed the same version and the same section. One may be describing the introductory unit while the other describes the advanced exercises.

If the source itself is inconsistent, preserve that conflict in the comparison. A majority vote among agents does not establish the underlying fact. Several workers can make the same error because they share a source, prompt, model, or assumption.

Independent approaches can sometimes help identify mistakes, but count that benefit through evaluation. Independence should describe the evidence and method, not merely the fact that several conversations exist.

Plan for failed and redundant workers

A worker may time out, return unsupported claims, or accidentally receive a duplicate assignment. The coordinator needs a response for each case.

Preserve useful partial results with their status. Reassign only the missing work. Cancel redundant assignments when the duplication is known. If one resource cannot be evaluated within the budget, the final report should say so rather than quietly ranking it from incomplete evidence.

Define whether a failed worker blocks the whole deliverable. A safety-critical comparison may require every assigned check. A preliminary reading list may be useful with one clearly labeled gap. The acceptance criteria should settle that choice before the failure occurs.

Use a small, reviewable experiment

For the learning center, select three short public resources and five questions. Save the exact source packet. Have one worker complete all reviews, then run a second condition with one resource per worker and a designated synthesizer.

Use the same rubric and blind the reviewer to the condition when practical. Record the final factual accuracy, evidence coverage, elapsed time, total usage, and review minutes. Deliberately duplicate one assignment and simulate one failed worker in a separate recovery exercise.

This article provides the experimental design. No live multi-agent comparison was run during preparation, so the table above remains illustrative.

A reusable prompt

Split this task only where the work is independent. Give each worker a nonoverlapping scope, fixed evidence requirements, budget, and stopping rule. Assign one owner to the final synthesis. Define how to handle conflicting claims, duplicate assignments, and failed workers. Compare the team with a single-worker baseline using accepted output quality, elapsed time, total cost, and review effort.

For students: collaboration is a skill to measure

Student teams can perform the comparison with people, AI workers where permitted, or a mixture. Each participant should understand the assigned material and retain its evidence. A final editor must reconcile differences rather than simply concatenate answers.

When a course assesses individual understanding, separate the collaborative research record from each student's own explanation. Disclose AI use according to the assignment rules. Parallel work should not make it impossible to tell who checked the sources.

Practice: justify one delegation

Choose a small comparison task and identify its independent parts. Write assignment contracts and a synthesis rubric. Estimate the sequential portion, then run or plan a controlled baseline comparison. Label estimates and measured results distinctly.

Completion check: Delegation has a stated purpose, evidence remains traceable, and any claimed benefit includes the cost of coordination and final review.

Stretch: Introduce one failed worker and one duplicate assignment. Recover the missing work without discarding accepted results or counting duplicated evidence as independent support.