Coordination and cost

Multi-agent runs can consume many times the tokens of a chat: orchestrator tokens plus every worker's exploration. Fan-out of 5 workers that each take 20k tokens is 100k+ before synthesis. Parallelism buys wall-clock. It does not magically buy efficiency.

The big idea

Budget the run before you celebrate the architecture:

cost ≈ orch_tokens × price_orch
     + Σ_workers (worker_tokens × price_worker)
     + synthesis_tokens × price_orch

Cap workers per request, cap tokens per worker, and abort low-value branches early. Start with 2 to 4 workers. Beyond that, plan quality and merge quality usually degrade unless the tasks are trivially disjoint, such as "embed these 20 independent PDFs." The nth worker adds tokens linearly and insight sublinearly.

Adaptive fan-out helps: start two workers, read partial results, spawn a third only if a gap remains.

Also budget engineering time. Distributed tracing across four agents, merge evals and dedupe caches is real work. If a single agent with hybrid search already meets the bar, spend the team time on retrieval and evals instead.

Walking through the failure modes

Duplicated work. Two workers scrape the same site. Fix: the orchestrator assigns disjoint scopes. Optionally share a cache of fetched URLs.

Conflicting edits. Two coding agents touch the same file. Fix: single-writer ownership, file locks, or serialize edits through the orchestrator. Give each worker a branch, a directory or a ticket id.

Lost information. A worker found a critical caveat; the summary dropped it. Fix: required fields for risks and citations, plus spot-check evals that score whether known landmines survive the merge.

Orchestrator bottlenecks. Everything waits on a slow planner. Fix: coarser tasks, async completion, or a workflow for the stable parts.

Herding. Workers share too much intermediate state and converge on the same wrong assumption. Fix: independence until merge. Do not over-share.

Make cost visible: tokens per run, workers spawned, cache hit rate on fetches. Teams that cannot see the multiplier cannot regulate it.

Common mistakes

• Fan-out as a default. Token cost shows up immediately. "We might need it later" is not a reason.
• No owner for a file. Two writers will collide.
• Sharing every intermediate note. Workers then herd onto one wrong assumption.

Check yourself

• Five workers at 20k tokens each, before synthesis: how many tokens is that? 100k, plus whatever the orchestrator and the merge use.
• Two coding agents edit the same file. What is the first fix? Give each worker its own branch or serialize writes through the orchestrator.

Key takeaways

• Cost is orchestrator plus every worker plus synthesis. Parallelism is a wall-clock trade, not a discount.
• Start with 2 to 4 workers. Cap tokens. Abort weak branches.
• Watch duplication, conflicting edits, lost caveats, planner bottlenecks and herding.

Further learning

Next, measure whether the merge is honest in evaluating multi-agent systems.