Single Agent vs Multi-Agent: What Anthropic's Subagent Guidance Means

Short answer

For most tasks that share one context, a single agent that loads skills and tools when it needs them now beats a multi-agent setup on quality, and often on cost and speed. That is Anthropic's finding in its October 2026 guide, Building agents: When to fold subagents into your main agent loop. Multi-agent still wins in three places: when one loop degrades from context bloat, when tasks run in parallel with fresh context, and when one agent picks the wrong tools or holds conflicting instructions. Split agents by context, not by org chart.

What Anthropic said

Anthropic's Applied AI team compared designs across several enterprise deployments. In its words, "a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality, and often at a lower cost and latency per task."

The guide gives two concrete cases:

• One company's customer-facing agent had a dozen specialist subagents. That made a consistent, on-brand experience hard to keep and added latency.
• One customer that tested its multi-agent harness against a single agent loop saw roughly 2x faster runtime with the single loop.

The proposed fix is progressive disclosure. The agent sees short descriptions of its skills and tools and loads the full version only when needed. Procedures move into skills. Actions and hard limits, such as validation and approval gates, move into tools. Anthropic also published a single agent loop kit that scores the system on the same eval before and after each change, so a subagent only goes away when the numbers say its replacement holds.

The post took off on X. One summary of the guide had passed 890,000 views and 5,600 bookmarks by 26 October 2026, two days after it went up. The Decoder's write-up ties it to a Vals study that found agent teams can be faster than single agents but do not deliver higher quality.

Why handoffs cost you

Every handoff between agents is a summary, and summaries drop detail. Anthropic calls the result a "telephone game": when agents are split by problem type, information passes back and forth and loses fidelity at each step.

The guide names the failure modes that show up in large organisations:

• Subagent sprawl. Every team ships its own agent, and nobody owns the overlap.
• The Frankenprompt. Every team adds a sentence to the system prompt, and the rules start to conflict.
• Local evals. Each team measures its own subtask, and nobody measures the whole.
• Latency. Each nested model loop adds time.

Anthropic's diagnosis is Conway's law: the agent architecture becomes a copy of the org chart.

When multi-agent still wins

Anthropic does not say subagents are dead. It says it uses them in its own agents, including its multi-agent research system, and that they are effective at "verification, parallel work, and modular permissioning." The guide keeps a subagent when:

1. A single agent loop shows measurable degradation from context bloat.

2. Several tasks need to run in parallel with fresh context windows that can be thrown away afterwards.

3. One agent struggles to choose the right tools, or the system prompt holds conflicting instructions.

If you do split, Anthropic's rule is to divide by context. Good splits are independent research paths, separate components with clean interfaces, and black-box verification. Each piece needs little from the rest of the system and hands back a compact result.

The guide also warns against merging blindly. In a sample support agent, folding one subagent into another 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, and the break only showed up in one run out of two. Measure before and after every change.

Single agent vs multi-agent at a glance

Anthropic adds one more limit for the single-agent side: tool choice tends to degrade past 30 to 50 tool definitions, so tools should be grouped by user intent rather than one per API.

What this means for a squad of AI agents

A squad of agents does not have to be the architecture Anthropic is warning about. The test is whether it is split by context, with bounded work and a check at the end, or split to mirror an org chart.

In xSquads for Work, each knowledge worker gets their own squad of specialised agents to delegate to and review. The squad works a shared task board on your own files and folders, and each bot claims the cards that match its role. Each card is a bounded piece of work, not a mid-conversation handoff. Finished work waits in Review until you approve it. It runs on Mac now, with Windows coming soon, and it is free through the waitlist.

The Software Engineering Squad shows the split. Blueprint writes the spec, Forge builds test-first, Check reviews the change, and Shield looks for security risks. Each agent has a named role and loads specific engineering skills. The builder and the reviewer are different agents, which lines up with one of the cases where Anthropic says a separate agent earns its place: verification.

Two things to take from the guide if you run a squad:

• Do not add an agent for every team or topic. Remove the agents a mission does not need. Anthropic's warning is about architectures that copy the org chart, not about having roles.
• Keep the check that catches regressions. Anthropic's sample support agent broke in one run out of two after a merge. A person reading what lands in Review is a check that can catch that kind of break.

The point of a squad is to make the person running it 10x more productive, not to run the most agents. You delegate, the squad works, and you review.

The same rule in a managed software factory

An AI software factory is a production line for code: a request comes in, agents do a bounded step, checks run, and a person decides what merges. A managed software factory is that line run for you, instead of a toolkit you assemble and staff yourself. On this site, that offer is xSquad operating as your code factory: agents open pull requests, and a person reviews before anything merges.

Anthropic's guidance fits that line. Each step is bounded and hands back something a person can read, a pull request, rather than a lossy summary passed to the next agent. The review sits with a person, separate from the agent that wrote the code.

Frequently Asked Questions

What is the difference between a single agent and a multi-agent system?

A single agent runs one loop with one context window and loads skills and tools as it needs them. A multi-agent system splits the work across several agents, usually an orchestrator and specialist subagents, each with its own context, instructions, and tools. The trade-off is shared context against isolation.

Is a single agent better than multi-agent?

Often, for work that shares one context. Anthropic found that a single agent with skills frequently beat subagent designs on quality across several enterprise deployments, often at lower cost and latency, and one customer saw roughly 2x faster runtime. It is not a universal rule: Anthropic says to measure each change with an eval.

When should you use subagents?

Use subagents when one loop degrades from context bloat, when tasks need to run in parallel with fresh context, or when one agent struggles to pick the right tools. Anthropic also names verification and modular permissions as strengths. Split by context, so each subagent needs little from the rest and hands back a compact result.

What is the difference between skills and subagents?

A skill teaches an agent how a team does a piece of work: the steps, rules, and judgement calls. The agent loads it into its own context when needed. A subagent is a separate agent with its own context window, instructions, and tools that hands a result back. Anthropic's guide suggests turning most subagent prompts into skills and their API calls into tools.

Does Anthropic's guidance mean AI agent squads are a mistake?

Not if the squad is split by context and reviewed. Anthropic warns against one agent per team or problem type. A squad in xSquads for Work gives each bot a named role on a shared board, each card is a bounded task, and finished work waits in Review for a person. Remove the agents a mission does not need.

How does a managed software factory use multiple agents?

On this site, xSquad runs the line as a code factory: agents open pull requests, and a person reviews before anything merges. Each step is bounded, and the hand-back is a pull request a person can read rather than a summary passed to another agent.

Related Reading

• AI agent workflow that doesn't fry your brain
• Agent Swarms: The Future of Software Development Teams
• What a managed AI software factory does
• What Is a Software Factory? Code and Dark Factory
• What Is a Personal AI Agent? From Chat Assistant to AI Teammates

Raina Rufus, Founder of xSquads

xSquads for Work is free. Join the waitlist and we'll send the download link.