How Many Parallel Agents Can Review Keep Up With?

Replay task durations and review demand through different worker counts. Pick the smallest pool that achieves the fastest reviewed finish within your spend cap.

Measure when finished work reaches the reviewer

Anthropic’s research system distributes exploration across agents with separate contexts and compresses their findings for the lead. Apply that idea to a mixed-duration pilot. Ask each worker to return a claim, its supporting evidence, and a reproducible check. Record when the packet becomes reviewable and how many minutes it takes a reviewer to accept it. Keep the packet shape fixed across runs with one worker and with several workers. Compare time to first accepted packet, time to last accepted packet, and peak waiting queue. For example, if six workers finish at minute 10 and each packet takes four minutes to review, the last packet clears at minute 34. That finish time tells you more about delivery than the worker completion time does.

How we built our multi-agent research system \ Anthropic

Time the review with isolated outputs

Choose a pilot where workers produce changes against the same baseline and send artifacts to one reviewed destination. Anthropic’s deployment guide describes using an overlay filesystem to keep changes in a separately inspectable layer. For a coding pilot, give each worker its own writable layer and record the baseline revision on every submission. Review each patch against that revision, then apply accepted changes in order to a disposable integration checkout and run the same acceptance test after each application. A patch that passes on its original baseline but fails after an earlier patch adds integration work. Attribute the diagnosis and retest minutes to its review duration. That turns merge friction into measured queue demand and shows when related tasks should share an owner or run in sequence.

Securely deploying AI agents - Claude Code Docs

Decide what can start at the same time

Anthropic’s Research system uses a planning agent and parallel search agents. Apply that pattern when a coordinator can give each worker a bounded assignment with inputs already available. For a report, give each worker a different source set and require a short claim list with evidence. For a coding task, agree on any shared interface before dispatching workers that will rely on it.

Ask each worker to return its artifact, how to verify it, and which assumptions could change the answer. If two workers need to invent the same interface or wait for the same unresolved decision, settle that decision first. The parallel run is ready when each assignment can start immediately and a reviewer can evaluate each result on its own.

How we built our multi-agent research system \ Anthropic

Make shared state explicit before launch

Git documents that linked worktrees share repository configuration by default. Give each writing agent its own checkout and branch, then assign separate output directories, test databases, or ports wherever the tasks would otherwise compete for a resource. Review repo-wide configuration changes centrally. If the job needs a per-worktree Git setting, verify Git version compatibility before enabling the worktree configuration extension.

In an illustrative job, one agent changes a CSS print layout and another updates a CSV formatter. Both start from the same approved commit and return one branch and a test result. If both suites write to the same fixture output directory, give them separate destinations before launch. Confirm that each branch changes only its assigned files, then run the full suite against the integrated change.

Git - git-worktree Documentation

Give the reviewer control over publication

Anthropic’s secure deployment guidance describes an overlay filesystem that stores an agent’s changes in an inspectable layer. Where that setup fits, give each agent its own writable layer over the input. For a coding run, you can also collect a branch diff from each worker. Keep publication credentials with the reviewer and require each submission to include the result, its verification output, and unresolved assumptions.

Review each submission against its assignment and rerun the acceptance tests on the combined output before publishing. Two patches can pass independently and still disagree about expected behavior. If the integrated test fails, identify the conflicting assumptions and assign the correction to one owner. Repeat the combined test before accepting the batch.

Securely deploying AI agents - Claude Code Docs

Count review time in the parallelism decision

Anthropic’s agent design guidance describes simple patterns implemented directly with model APIs. Start with one operator launching workers and collecting results so the first trial measures the coordination the job actually needs. Compare sequential and parallel runs using the same acceptance checks. Record elapsed time through approval, model and runtime charges, and time spent reviewing, retrying, or resolving conflicts.

In an illustrative comparison, three independent ten-minute tasks followed by five minutes of review take 35 minutes sequentially or 15 minutes if all three start together and still need five minutes of review. If parallel review and rework grow to 25 minutes, the elapsed-time advantage disappears. For competing solutions, count the spend on both attempts and choose between them using an agreed test.

Building effective AI agents \ Anthropic

Sources