AI Agents Engineering · Orchestration & Multi-Agent · Lesson 1 of 3
Single vs multi-agent
Multi-agent systems split work across multiple LLM loops. The 2023–25 'agent swarm' debate has settled: peer-to-peer agent debate died in production, and the pattern that survived is one orchestrator that owns the full context, plus ephemeral sub-agents used as context firewalls — each does isolated work and returns a compressed summary.
- Decompose by context boundaries, not job titles: spawn a sub-agent when work would pollute or overflow the main context, or genuinely parallelizes (read-heavy research fan-out) — not because 'planner/coder/critic' sounds tidy.
- Single-writer principle: extra agents contribute intelligence (research, review, critique); exactly one agent owns actions/writes. The verification sub-agent is the most reliably useful pattern.
- Budget honestly: multi-agent typically costs 3–10× the tokens, and at equal token budgets a single agent often matches it. Watch bad handoffs, context loss, coordination loops.
Multi-agent isn't freeEach added agent multiplies cost, latency, and coordination bugs. Often a single well-tooled agent (or a plain workflow) beats an elaborate swarm — token budget, not architecture, explains most of the performance difference.
◆ Lock it in
- The settled 2026 shape: orchestrator owns context; ephemeral sub-agents as context firewalls returning compressed summaries.
- Decompose by context boundaries, not problem type; one single writer owns actions; verification sub-agents are the reliable win.
- Multi-agent costs 3–10× tokens and multiplies coordination failures — justify it.
Feynman drill — say it out loudGive one task that justifies multi-agent and one that a single agent handles better, and say why.
Step 1 rate your confidence · Step 2 pick your answer
The biggest risk when moving from a single agent to a multi-agent system is:
Step 1 — how sure are you?