Multi-agent handoff budgets (1-2k return summary vs shared file store)

“Multi-agent” systems split a big job across several AI agents: a lead hands out subtasks, workers go off and explore, and results flow back. The quiet make-or-break question is the handoff. What exactly does each agent tell the other, and how long should that message be?

Three lanes

Down, from lead to worker. The lead writes a brief. Anthropic’s Research team found that short briefs (“research the semiconductor shortage”) made workers misread the scope or run the same searches as their siblings. Their fix was a four-part brief: objective, output format, which tools and sources to use, and boundaries. They also wrote effort rules into the lead’s instructions: a fact lookup gets 1 agent and 3 to 10 tool calls; a comparison gets 2 to 4 workers; complex research can get 10 or more with clearly split responsibilities.

Up, from worker to lead. A worker may spend tens of thousands of tokens exploring, then returns a condensed summary, often 1,000 to 2,000 tokens, per Anthropic’s context-engineering post. Claude Code’s subagents work the same way: logs and file reads stay with the worker, only the summary comes back.

Sideways, into files. Anything bulky goes into a file, and the worker passes back a path. That avoids a game of telephone: the original stays exact, and the lead can open it if a summary looks thin.

What earns a spot in the 1 to 2k return:

the answer to the question it was given
decisions it made that someone else could contradict
what it tried that failed, so nobody repeats it
open questions plus a status (done, partial, blocked)
paths or IDs for everything else

Why the budget matters

Tokens. Anthropic measured agents at roughly 4x the tokens of a chat, and multi-agent systems at roughly 15x. On the BrowseComp benchmark, token usage alone explained 80% of performance differences. Worked example: ten workers returning 1.5k each add 15k tokens to the lead. Ten raw 40k transcripts would add 400k and blow past most windows.

Reliability. Google Research and MIT ran 260 configurations in “Towards a Science of Scaling Agent Systems” (December 2025). Independent parallel agents amplified errors in the trace 17.2x versus a single agent; a central orchestrator that reviews worker output held that to 4.4x. Every multi-agent variant hurt step-by-step planning, by 39% to 70%. Centralized coordination lifted structured financial reasoning by 80.8%. And once a single agent already scored above about 45%, adding agents gave negative returns. The MAST paper, studying 1,600+ traces across 7 frameworks, found 14 failure modes, and a whole category of them happen at handoffs.

The debate

Cognition’s Walden Yan wrote “Don’t Build Multi-Agents” in June 2025: share full traces, because actions carry hidden decisions and conflicting decisions produce bad results. Anthropic’s Research system is the counterweight: an Opus 4 lead with Sonnet 4 workers beat single-agent Opus 4 by 90.2% on their internal research eval, though the same post says most coding tasks have fewer truly parallel pieces.

Read closely, the camps land near each other. Parallel workers do well on read-heavy breadth work and return capped summaries. Writing and design calls stay in one thread or go through a reviewer. And measure the single-agent baseline before adding agents.

An analogy: the newsroom

An editor sends five reporters out on a story. Nobody wants each reporter to read their entire notebook aloud back at the office. They want a tight memo: here’s what I found, here’s who I talked to, here’s what fell through, here’s where my full notes are. The editor writes the article. Five people writing the same article at once would be chaos.

Try this

Run a tiny multi-agent experiment by hand. Open three separate chats. In two of them, give each a research subtask (“summarize the pros of electric cars,” “summarize the cons”). Ask each to finish with a handoff of under 150 words: findings, uncertainties, and what it didn’t check. Paste only those two handoffs into the third chat and ask for a balanced recommendation. Then try it again pasting the full answers instead. Which final answer was clearer?

Use it at dinner

“When AI agents work as a team, the best ones hand off like reporters to an editor: a short memo plus where the full notes live.”