Chapter Ten · Multi-Agent Systems

Multi-Agent Systems

Six topics on splitting an agent, and on the far more common case for not splitting it: the two circumstances where a second agent earns its keep, why a subagent's own context window is the one argument that survives scrutiny, what belongs in a task envelope and what has to come back, the two channels a handoff travels on, the difference between parallel tool calls that are nearly free and parallel agents that cost 3.2 times the tokens, and the split Sundry actually shipped — reported with the number it made worse.

6 topics

The Sundry agent now carries nine tools, a policy retriever, a five-state returns machine and a system prompt that has grown three sections. It is the biggest single thing in the codebase, and the obvious next move — the one every architecture diagram on the internet agrees on — is to split it into a team. A researcher, a writer, a critic. Boxes with arrows between them.

This chapter is about how often that is the wrong move. A second agent is not a refactor: it is a new failure surface, and the failures it adds are specific rather than theoretical. Information crosses the boundary as a summary and the evidence behind it does not. A subagent that ran out of turns reports a tidy conclusion anyway. Two agents write to the same order. None of that shows up in the diagram, and all of it shows up in the eval report — which is why the decision belongs to Chapter 9's instrument rather than to anybody's taste.

Two cases do win, and they are worth learning to recognize: a context that must not contain something, and work that is genuinely independent with a deadline attached. Sundry's own split is built on the first. It is a cheap triage agent that answers the simple majority and routes the rest to two specialists, and it moved handle time from 34 seconds to 19 while raising cost per ticket by 11%. The resolution number went up by two points, which on a 120-case set means nothing at all. Topic 58 reports all three of those honestly, along with the agents that were measured and rejected and the conditions under which Sundry would merge the whole thing back into one loop.

What each shape keeps, and what the second one gives away at every boundary
One agent, nine toolsthe default answer
Keeps: every fact the run gathered is visible to the decision that uses it, in its original wording. One message list, so a bug reproduces by replaying it. One place counting turns, one measuring spend, one dispatcher deciding what may execute. Loses: nothing to a handoff — but the context fills up, and tool selection falls as the list grows.
Orchestrator plus specialistsa new failure surface
Keeps: contexts small, tool lists narrow, and the simple majority never loads the machinery that moves money. Sundry's split took handle time from 34 seconds to 19. Loses: the evidence behind every crossing — the receiver gets a summary, a child that ran out of turns still reports a tidy conclusion, two agents can write the same order, and the bill goes up 11% per ticket.

Topics in This Chapter

Topic 53
When One Agent Is Not Enough
Most multi-agent designs are one agent with extra latency: a role-played team of researcher, writer and critic scored 84% against the single agent's 86%, at 2.3 times the tokens and nearly twice the handle time. The two splits that genuinely win, what one handoff costs in tokens and turns, and how to decide with the five-number eval report instead of a diagram.
Architecture
Topic 54
Context Isolation Is the Real Reason
A subagent reads forty carrier records and returns three sentences, keeping 12,000 tokens of noise out of a decision context budgeted at 8,000. The summary is the interface, so it carries required fields, a completion status and a confidence signal — and the failure to name is the tidy summary from a subagent that ran out of turns.
Isolation
Topic 55
Orchestrator and Subagents
The task envelope is constructed from state, never copied from the transcript: the question, the facts, the constraints, the tool list, and a budget carved out of the parent's twelve turns. Write authority stays with the parent, budgets divide rather than duplicate, and depth is capped at one level after a ticket that cost $2.30.
Delegation
Topic 56
Handoffs and Shared State
Structured state travels through the store both agents read; the message carries the summary and the ask, never the transcript. The completed-actions list is what stops a receiver repeating a write the sender already made, and an ownership rule written into the same transaction is what stops two agents passing one ticket back and forth.
Handoffs
Topic 57
Parallelism and Its Bill
Three independent read-only calls issued in one turn take 2.0 seconds instead of 2.9 and bill exactly the same tokens — the cheapest latency win in the book, routinely left on the table. Three agents in parallel take 40% off the wall clock and multiply tokens by 3.2, which is a different decision needing a different justification.
Parallelism
Topic 58
Sundry: Triage and Specialists
Handle time 34 seconds to 19, resolution 86% to 88%, cost per ticket up 11% — and the honest reading, which is that a two-point move on a 120-case set is not evidence of anything. What broke on multi-intent tickets, which two agents were measured and rejected, and the written conditions under which the split gets undone.
Case Study