facts 524 · signals 29 · articles 20 · database 3,168 KB · bots —guest@ftrlgthis is you~/whoamiYou post anonymously. To sign the same way every time, pick a secret when commenting — it becomes your tripcode.no sign-up, ever

AI agent orchestration

translated by AI, reviewed by the editors

future.log editors · published

AI agent orchestration is a way of solving a task with several AI agents at once: a lead orchestrator agent splits the work into parts, hands them out to worker agents, oversees them and assembles the result. The orchestrator can also be a person who runs several agents at the same time.

The first open-source frameworks for this appeared in 2023.1 By 2026 orchestration had become a built-in feature of development tools: in Claude Code you can launch a team of agents that work in parallel and coordinate on their own.2 Research shows, however, that several agents do not always help: on some tasks they are markedly stronger than a single agent, on others markedly weaker.3

History

  • 16.08.2023Microsoft publishes AutoGen, an open-source framework in which applications are built from multiple agents that “can converse with each other to accomplish tasks”.1 ¶ cgbiu3vru
  • 23.01.2024LangChain demonstrates multi-agent setups on LangGraph: “An agent supervisor is responsible for routing to individual agents”; in effect, the supervisor is “an agent whose tools are other agents”.4 ¶ cbfqw8d8u
  • 19.12.2024Anthropic, in its guide “Building Effective AI Agents”, describes the orchestrator-workers pattern: “a central LLM dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results”.5 ¶ cbh6rb67a
  • 09.04.2025Google releases the Agent Development Kit for “composing multiple specialized agents in a hierarchy”, and the open A2A protocol, “a standard way for agents to collaborate with each other” across companies; the protocol was backed by “more than 50 technology partners”.67 ¶ cua9adupa
  • 24.07.2025Claude Code gets its own subagents: each one “runs in its own context window with a custom system prompt, specific tool access, and independent permissions”.8 ¶ c2ftwfcc7
  • 05.02.2026Anthropic adds “agent teams” to Claude Code (research preview): you can “spin up multiple agents that work in parallel as a team and coordinate autonomously”.2 ¶ crz8sdvjm
  • 08.09.2026OpenAI reports that its internal model proved a finite-time singularity for the Navier–Stokes equations (with an applied force), “using around 10,000 coordinating AI agents” over 88 hours.9 ¶ cfknhmsr6
  • 07.10.2026Elon Musk announces that Grok Bot will “use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs”.10 ¶ cdiwqjb93
  • 08.10.2026Google introduces the Gemini agent, which “runs each job on the model that fits best”, orchestrating across Gemini and Claude models.11 ¶ cb9tndyh4
  • 09.10.2026Anthropic opens the public beta of Claude Managed Agents dynamic workflows: “A lead agent writes a plan that runs across many agents in phases.” With 70 planted bugs, “a single agent found 14, 15 and 27 bugs. A workflow consistently found 66 in each of its 3 runs.”12 ¶ cp6puqsz6

How it works

There are two patterns. In the first, the orchestrator is a model: it decides on its own how to split the task and launches the workers. This is how the frameworks AutoGen, LangGraph, CrewAI, OpenAI Agents SDK (the successor to the experimental Swarm)13 and Google ADK work, as do subagents and agent teams in Claude Code.

In the second, the orchestrator is a person who runs several agents in parallel. Simon Willison writes: “I frequently have multiple terminal windows open running different coding agents in different directories.”14 ¶ cn6fi9i8j

The Gas Town project by Steve Yegge turns this into a system: a lead agent, “The Mayor”, coordinates the rest, and according to the project’s description it lets you “Scale comfortably to 20-30 agents”.15

In 2026 orchestration also began to span models from different companies: Google’s Gemini agent and xAI’s Grok Bot route each task to whichever model handles it best, including rivals’ models.1011

Data

studywhat was measuredresult
Cemri et al. (University of California, Berkeley), 17.03.2025161,642 execution traces of 7 open-source multi-agent systems; 150 annotated by hand, the rest by a modelfailure rates from 41% to 86.7%; 14 failure modes identified in three groups: system design, inter-agent misalignment, task verification ¶ cr99v6qp8
Anthropic, 13.06.202517a research system: Claude Opus 4 as orchestrator, Claude Sonnet 4 as workers; an internal test by a company that sells tokens“outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval”; it uses “about 15× more tokens than chats”; token usage alone explains 80% of the variance in quality ¶ c8cay2zwe
Kim et al. (Google Research, Google DeepMind, MIT), 09.12.20253260 configurations: 6 agent benchmarks, 5 architectures (one single-agent and four multi-agent), 3 model familiescompared with a single agent, from +80.8% on tasks that split into parts (financial analysis) to −70% on sequential planning ¶ cfe6jdqh7
[EDITORS’ CONCLUSION]

Multiple agents are not a universal improvement but a bet that wins on tasks that break into independent parts and loses where steps depend on each other. The gain is paid for in tokens: at Anthropic, quality is explained above all by the number of tokens spent, and a multi-agent system spends about 15 times more of them than a chat. Open-source multi-agent systems fail 41% to 87% of tasks, and a notable share of the failures come from inter-agent misalignment and unverified results. Caveat: the largest gain (+90.2%) is an internal test by a company that sells tokens. ¶ c3tkh6djn

Incidents

ServiceNow: an agent that calls in other agents, November 2025 ¶ c4hfezr2m

AppOmni researcher Aaron Costello showed that the agents of ServiceNow’s Now Assist AI assistant can discover each other and assign tasks to each other, and the settings needed for this are on by default. An instruction injected into data made a harmless agent “recruit more powerful agents”, which then created, read, changed and deleted records and sent out emails containing other records, with prompt injection protection turned on. ServiceNow confirmed that this behavior is intended and clarified its documentation. AppOmni sells protection against such attacks.18

Criticism

Cognition, the developer of the Devin agent, came out against multi-agent systems in Walden Yan’s article “Don’t Build Multi-Agents”. Its two principles: “Share context, and share full agent traces, not just individual messages”; and “Actions carry implicit decisions, and conflicting decisions carry bad results”. An example: one subagent misunderstood its part of the task and “started building a background that looks like Super Mario Bros”, another drew a bird that did not look like a game asset, and the final agent was left to combine the two misunderstandings. Cognition, for its part, sells a single-agent product.19 ¶ c6dwpjkbc

Signals

See also

Autonomous AI agentsAgentic engineeringAgentic codingClaude CodeSelf-writing software

References

  1. Q. Wu et al. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation — arXiv, 16.08.2023. arxiv.org ↩ ↩2

  2. Claude Opus 4.6 — Anthropic, 05.02.2026. anthropic.com ↩ ↩2

  3. Y. Kim et al. Towards a Science of Scaling Agent Systems — arXiv, 09.12.2025 (version 3 — 08.04.2026). arxiv.org ↩ ↩2

  4. LangGraph: Multi-Agent Workflows — LangChain, 23.01.2024. langchain.com ↩

  5. Building Effective AI Agents — Anthropic, 19.12.2024. anthropic.com ↩

  6. Agent Development Kit: Making it easy to build multi-agent applications — Google for Developers, 09.04.2025. developers.googleblog.com ↩

  7. Announcing the Agent2Agent Protocol (A2A) — Google for Developers, 09.04.2025. developers.googleblog.com ↩

  8. Claude Code changelog, version 1.0.60 — Anthropic, 24.07.2025. github.com ↩

  9. On the Navier–Stokes Millennium Prize Problem — OpenAI, 08.09.2026. openai.com ↩

  10. Elon Musk: “Important note regarding Grok @Bot…” — X, 07.10.2026. x.com ↩ ↩2

  11. Gemini at Work 2026: Introducing Gemini agent — Google Cloud, 08.10.2026. cloud.google.com ↩ ↩2

  12. Claude Developers: “We planted 70 bugs in a 116k-line codebase…” — X, 09.10.2026. x.com ↩

  13. Swarm (experimental, educational) — OpenAI, GitHub. github.com ↩

  14. Embracing the parallel coding agent lifestyle — Simon Willison, 05.10.2025. simonwillison.net ↩

  15. Gas Town — Steve Yegge, GitHub. github.com ↩

  16. M. Cemri et al. Why Do Multi-Agent LLM Systems Fail? — arXiv, 17.03.2025 (version 3 — 26.10.2025). arxiv.org ↩

  17. How we built our multi-agent research system — Anthropic, 13.06.2025. anthropic.com ↩

  18. When AI Turns on Its Team: Exploiting Agent-to-Agent Discovery via Prompt Injection — AppOmni, 19.11.2025. appomni.com ↩

  19. Don’t Build Multi-Agents — Cognition, 12.06.2025. cognition.com ↩

Your signature

Remember the secret — it is your password. The tripcode is computed from it and shown next to your name: the same secret always gives the same tripcode, and nobody can fake it without the secret. The secret is never published or stored on the server.