Why This Matters
Some workloads are embarrassingly parallel: 50 files to review, 100 tickets to triage, a test suite to generate function-by-function across thirty modules. Walking them serially with one agent takes thirty minutes — most of it waiting on the model, not thinking. Three parallel agents finish in a third of the wall-clock time. Five finish in a fifth. The math is irresistible, but the wiring is treacherous: naive fan-out lands you in merge hell, race conditions, or a 5x cost spiral. Parallel agent supervision is the narrow case where multi-agent actually earns its keep. This guide is the wiring diagram — three preconditions that must all hold, the supervisor-as-code pattern, deterministic merge rules, the concurrency budget, and the four failure modes (divergent agent, resource race, supervisor over-trust, latency cascade) with their fixes.
The Problem
You have 50 files to review. Or 100 tickets to triage. Or a test suite to generate function-by-function across thirty modules. A single agent walks them serially and takes thirty minutes — most of which is waiting on the model, not thinking.
The shape of the work is screaming at you: this is embarrassingly parallel. Three agents working slices of ten files each finish in roughly a third of the wall-clock time. Five agents finish in a fifth. The math is irresistible.
It's also a trap, if you wire it up naively. Teams who reach for multi-agent without the right wiring end up in merge hell, ping-pong loops, and a 5x cost spiral instead of a 3-5x speedup. Parallel agent supervision is the narrow case where the math actually works. This guide is the diagram for that wiring.
The Core Insight
Parallel agents work when three conditions hold at the same time. Miss any one and you are back in the anti-patterns catalog.
- The work partitions cleanly with no shared state. Each agent gets a slice no other agent will touch.
- The supervisor merges deterministically. The merge step is a rule, not a model call. Alphabetical order. Line numbers. Concatenation by key.
- Failures don't cascade. One agent timing out or going off-spec does not corrupt the other four.
The default advice for multi-agent is "don't." This guide is the exception. The exception is narrow on purpose.