Frontiers · Long-horizon autonomy
You can describe the dominant 2026 multi-agent pattern — one orchestrator coordinating N specialist sub-agents — and the tradeoffs that come with using it instead of a single agent.
The orchestration UI from 6.7.2 sits on top of a question: under the hood, is it one agent doing the work in a loop, or is it many agents coordinating? By 2026, the production answer is increasingly "many."
Anthropic's multi-agent research system post is the canonical reference: one orchestrator decomposes a goal, dispatches sub-agents to do specialist work in parallel, and synthesizes the outputs. On Anthropic's own internal research evaluation, the pattern delivered roughly a 90% performance gain over single-agent — at roughly 15x the token cost. That cost ratio explains why multi-agent has not eaten everything, and why this chapter spends as much time on when not to as on how.
The framework landscape consolidated fast. LangGraph won the production-default slot — graph topology maps cleanly onto auditing, rollback, and human-in-the-loop checkpoints. AutoGen, the Microsoft research stack of 2024, was folded into the unified Microsoft Agent Framework v1.0 GA in April 2026. OpenAI's Swarm was replaced by the OpenAI Agents SDK.
And the most important counterweight to all the announcements: Anthropic's AI Organizations Can Be More Effective but Less Aligned (April 2026) — a multi-agent system can outperform a single agent on capability and simultaneously drift further from the operator's intent. More throughput, more coordination risk.
This chapter walks three lessons:
Chapter contains 3 lessons.