Trend · AI agents

Stop building agent swarms: when one agent with skills wins, and when subagents still earn their place

Updated Oct 26, 2026 · By Mishal Brands Editorial

Checked Oct 26, 2026. Every claim on this page was checked against the sources listed at the bottom on that date.

The short answer

Anthropic's Applied AI team says that, across several enterprise deployments it compared, a single agent that loads skills and tools on demand frequently beat both a "one giant prompt" design and a multi-subagent design on quality, and often cost less and ran faster per task. Subagents still make sense in three cases: when one agent measurably degrades from context bloat, when tasks need to run in parallel in fresh context windows, or when one agent struggles to pick the right tool or has conflicting instructions. Anthropic's advice is to measure: fold in one subagent at a time and keep it only if the folded version scores worse.

What it is

On Oct 25, 2026, The Decoder covered new Anthropic guidance titled "Building agents: When to fold subagents into your main agent loop," published alongside an open-source migration kit on GitHub, anthropics/single-agent-loop-kit (Apache-2.0).

The core ideas, in Anthropic's words and as reported:

Subagents are agents with their own context window, tools and instructions, run by a coordinator. Anthropic says they're good at verification, parallel work and modular permissions, and it still uses them (for example in its multi-agent research system).
The problem: large companies often give each team its own subagent, so "the agent architecture becom[es] a copy of their org chart" (Anthropic calls this Conway's law at work). Anthropic lists the failure modes as lost context in handoffs, subagent sprawl, the "Frankenprompt" (every team adds a sentence to the system prompt), overly local evals, and added latency.
The alternative: progressive disclosure. The agent sees short descriptions of skills and tools and only loads them fully when needed (skills for know-how, tool search for tools).
Anthropic's own data points (claims from Anthropic):
One customer that compared its multi-agent harness with a single agent loop saw a roughly 2x faster runtime with the single loop.
Anthropic says it cut Claude Code's system prompt for more advanced models by over 80%, partly by progressively disclosing skills and tools.
Tool choice "tends to degrade" once you pass 30–50 tool definitions in the system prompt, so group tools by user intent, not one per API.
Anthropic credits model progress: it says that starting with Opus 4.5, models could manage far more in a single loop than most teams expected.

The Decoder also notes outside work pointing the same way: a Vals study found agent teams can be faster but don't deliver higher quality, and that scaling from 10 to 30 or 100 agents added cost without quality gains (as reported by The Decoder).

Plain-English checklist: one agent or subagents?

Question	If yes, lean toward...
Does your single agent measurably get worse as context grows?	Subagent (or call one as a tool)
Do several tasks need to run in parallel with throwaway context?	Subagents
Does one agent keep picking the wrong tool or hit conflicting rules?	Subagent, or first fix tool grouping and the prompt
Do two specialists usually run back-to-back or touch the same data?	Merge them
Is the split mostly "one subagent per team"?	Fold into skills + tools
Does a specialist need to talk to the user directly?	Full handoff is OK; otherwise keep the main agent in charge
Do you have no eval yet?	Build one first (Anthropic says even 10 golden examples is a start)

How to use it, step by step

This follows Anthropic's article and the kit's README.

Build a small end-to-end eval first. Anthropic says 10 golden examples is enough to start. Grade the final answer against a rubric; while you still have a router, track whether it picked the right subagent.
Try the kit's free demo. The repo ships examples/support-desk, a support agent with a router and six subagents. The README says this check needs no API key and spends nothing: bash cd single-agent-loop-kit/examples/support-desk KIT_MODEL=fake python3 evals/run_eval.py --variant smoke That run is a plumbing check, not a score. Real scores come from the runbook.
Copy the skill into your own agent repo: bash git clone https://github.com/anthropics/single-agent-loop-kit cp -r single-agent-loop-kit/skill <your repo>/.claude/skills/single-agent-loop
Open Claude Code in your repo and say: Run the single-agent-loop skill. Per the README, it audits first and gives each subagent a verdict: fold, merge, call as a tool, or keep.
Merge tightly coupled pairs. Anthropic suggests merging specialists that chain on each other's output, touch the same data, or are often called together.
Fold the rest in one at a time. Turn each subagent's prompt into a skill and its API calls into MCP tools, then rerun the eval.
Look at which tasks flipped, not just the score. Anthropic's own test showed why. Folding one subagent cut latency but broke a process that had been passing, and that break showed up in only one run out of two. Run each task more than once.
Keep the ones that earn their place, and call them as tools so the main agent keeps the conversation.
Set ownership. One small central team owns the system prompt and shared eval; each domain team owns its skills.

Note: the README describes the kit as a "Reference implementation. Not maintained and not accepting contributions." It stops for your approval after the audit, before any paid run, after the baseline and after each change, and shows a token estimate before paid runs.

Alternatives and how they compare

Approach	What it is	Best for	Watch out for
Single agent loop with skills + tool search (Anthropic's recommendation)	One model, one context window, loading skills and tools on demand	Most enterprise and assistant-style agents	Needs a disciplined system prompt and good skill/tool descriptions
One giant prompt	Everything in one system prompt, all tools loaded	Small, simple agents	Anthropic says the skills design often beat it on quality
Orchestrator + subagents	A coordinator hands work to specialists with their own context	Parallel research, verification, clean component boundaries	Lossy handoffs, latency, sprawl
Subagents called as tools	Main agent keeps the conversation and calls specialists like functions	A first migration step; keeping a few specialists	Each call is still a nested loop with a summary handoff
Generator + verifier	One agent does the work, a separate one checks it	Black-box checking (e.g. running tests)	Adds a step; keep the verifier narrow

Want to build agents without writing an orchestration layer? See how no-code automation tools compare: n8n, Make vs Zapier, and how to plan your first automation.

Pricing

single-agent-loop-kit: free and open source (Apache-2.0 license, per the repo). Checked Oct 26, 2026.
Running it: the smoke test is free (no API key). Real eval runs call a model and cost API tokens; the kit shows a token estimate before any paid run. We did not run the paid steps, so we don't quote a cost.
Claude plans and API pricing: see our Claude review for current pricing.

FAQ

Is Anthropic saying multi-agent systems are dead? No. Anthropic says subagents "still have their place," including at the research frontier, and it uses them itself. Its point is that many companies split work into subagents that the model no longer needs.

When should I keep a subagent? Anthropic lists three cases: measurable performance loss from context bloat, tasks that need to run in parallel with fresh context windows, or a single agent that struggles to choose tools or has conflicting instructions.

What's the difference between a skill and a subagent? A subagent is a separate agent with its own context window. A skill is know-how (steps, rules, judgment calls) that the main agent loads only when it needs it, so everything stays in one context.

How much faster is a single agent? Anthropic says one customer saw a roughly 2x faster runtime with a single agent loop than with its multi-agent harness. That's one comparison, not a general benchmark.

How many tools can one agent handle? Anthropic says tool choice tends to degrade past 30–50 tool definitions in the system prompt, and recommends grouping tools by user intent and using tool search with deferred loading.

Is the single-agent-loop-kit free? Yes, it's open source under Apache-2.0. The smoke test needs no API key; real eval runs use paid model calls.

Related on Mishal Brands

Sources (checked Oct 26, 2026)

Anthropic, "Building agents: When to fold subagents into your main agent loop" (no date shown on page): https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop
GitHub, anthropics/single-agent-loop-kit (README, Apache-2.0): https://github.com/anthropics/single-agent-loop-kit
The Decoder, Matthias Bastian, Oct 25, 2026: https://the-decoder.com/anthropic-says-one-ai-agent-beats-many-subagents-now-that-models-can-handle-complexity-alone/

Disclosure: this page has no affiliate links. No vendor paid for or reviewed this page. See our affiliate disclosure and how we review.