Anthropic says one AI agent beats a team of subagents. Here's when that's true

For the last year, the default way to build a serious AI agent was a team: an orchestrator up top, specialist subagents underneath, each with its own prompt. Anthropic now says that for most jobs, that's the wrong shape. Its new guidance argues that one agent, loading skills and tools only when it needs them, often does better on quality, cost and speed. (Anthropic; The Decoder, Oct 25, 2026)

What Anthropic actually said

The post, "Building agents: When to fold subagents into your main agent loop," makes three arguments. (Anthropic)

• Shared context gives better answers. An agent that has seen every earlier step can connect them. A set of summaries handed back by subagents loses detail at every handoff.
• Progressive disclosure fixes prompt bloat. The agent sees short descriptions of its skills and tools and loads the full version only when needed. Anthropic says this helped it cut Claude Code's system prompt for newer models by over 80%.
• Models can handle more now. Anthropic says that starting with Opus 4.5, models could manage far more in a single loop than most teams expected.

The Decoder reports that in one comparison, Anthropic found the single-agent setup ran about twice as fast as the multi-agent one. (The Decoder, Oct 25, 2026)

Anthropic's commerce-agent guide makes the same case in plainer terms: in its comparisons across several enterprise deployments, a single agent with skills "consistently has outperformed" both one giant prompt and the subagent design on quality. (Anthropic)

The "Frankenprompt" problem

Anthropic also names a failure pattern most teams will recognize. Agent architecture starts copying the org chart, with every team shipping its own subagent. Then every team adds rules to the shared system prompt until they contradict each other. The Decoder reports Anthropic's term for that: a "Frankenprompt." (The Decoder, Oct 25, 2026)

Its fix is organizational as much as technical. A small central team owns the system prompt and the shared eval, and each domain team owns its own skills. Anthropic also notes that tool selection tends to degrade beyond 30 to 50 tool definitions, so skill libraries need owners to stop them from sprawling. (The Decoder, Oct 25, 2026)

When subagents still win

This isn't "never use subagents." Anthropic keeps them for narrow, self-contained jobs that benefit from their own context window. Deep research is the common example: the subagent reads, runs code and hits dead ends, and only a compact answer comes back. A domain that already has its own purpose-built agent, such as one with its own compliance rules, can also take over the conversation through a hand-off. (Anthropic)

It also warns against merging blindly. In one test on a sample support agent, folding in a subagent lowered latency but broke a process that had been passing: the main agent answered on its own instead of asking the shipping specialist for the delivery date. (Anthropic)

How to test it on your own agent

Anthropic's open kit lays out the process in three steps. (GitHub: anthropics/single-agent-loop-kit)

1. Measure. Score your current system on an eval. If you don't have one, the kit helps you build one.
2. Combine. Merge subagents that always run together.
3. Fold. Turn each remaining subagent into a skill plus tools, one at a time, and compare. Keep the ones that earn their own context.

The kit's own advice if you have no subagents yet: start with one loop and add tools and skills only as testing shows you need them. (GitHub)

Want the short version? See our single agent vs subagents checklist, and our explainer on Pomerado.

Why it matters: If you're building agents, or buying tools built on them, "more agents" is no longer a selling point by itself. Ask whether the setup has been measured against a single agent with good skills.

Earlier on Mishal Brands: our first Weekly AI Brief covered Agent Skills moving to an open foundation. (Mishal Brands on X, Oct 26, 2026)

Other top stories

Oracle puts agents inside Fusion apps. At Oracle AI World in Las Vegas, Oracle announced Fusion Companion and Fusion Agentic Home, which let employees state a goal and have the system pick the right transactions, workflows, approvals and agents, including through supported AI tools such as ChatGPT. Oracle also said 25 new "Claw-powered" agentic applications are available across its ERP, HCM, supply chain and CX suites. (Oracle, Oct 26, 2026; Oracle, Oct 26, 2026)

Kimi K3.1 Pro keeps climbing. X News reports Moonshot's model reached #11 in Agent Arena, a week after we covered its launch. (X News, updated Oct 25, 2026; Mishal Brands, Oct 25, 2026)

New tools to try

1. Single Agent Loop Kit (Anthropic) An open reference kit for moving a multi-agent system toward one agent loop, with an eval-first process for deciding which subagents to keep. Anthropic says it isn't maintained or accepting contributions. Website: https://github.com/anthropics/single-agent-loop-kit

2. Pomerado A YC Fall 2026 startup that turns any website into a hosted API or MCP server your agent can call, instead of driving a browser. It launched Oct 25, is open source, and has a free tier. The company claims its integrations are 2-4x faster, 10-1000x cheaper and 30% more reliable than browser use (vendor-reported). It connects to Claude Code and Codex. Website: https://pomerado.ai · Source: Launch YC, Oct 25, 2026

3. Nanbeige4.3-33B (open weights) A 33B dense model from BOSS Zhipin's Nanbeige LLM Lab, built for agentic coding and tool use, with a 512K-token context window and an Apache 2.0 license. Weights went up Oct 24. On Nanbeige's own tests it scores 80.9 on Terminal-Bench 2.1 versus 73.0 for Qwen3.8-27B; Qwen stays ahead on SWE-Bench Pro. Website: https://huggingface.co/Nanbeige/Nanbeige4.3-33B · Sources: Data Studios, Oct 25, 2026; AI/TLDR

4. Oracle Fusion Companion For teams already on Oracle Fusion Cloud: describe the outcome you want and Fusion Companion coordinates the workflows, approvals and agents within your existing permissions. Website: https://www.oracle.com/news/announcement/ai-world-oracle-introduces-new-ai-powered-ways-to-execute-work-across-oracle-fusion-applications-2026-10-26/

Are you running one agent or a team of them? Reply on X.