One agent or several? Anthropic’s guide to skills for your business

In 60 seconds: If you are deciding between one AI agent and several, start with a complete task and one agent with reusable procedures, called skills, and tools with limited access. Add another agent when tests justify separate context, parallel work or access boundaries. Anthropic’s guidance can help you review an architecture you already bought; decide using results, cost per accepted task and observed errors. Below is a distribution example and an evaluation you can prepare in an afternoon.

Imagine a small distribution business in Latin America that paid for six agents, an orchestrator and a router. A quote now fails somewhere between sales and the warehouse. Nobody knows which detail went missing or what reaching that failure cost. This is a synthetic example, not a Kiia client story.

In its October 22, 2026 guide, Anthropic proposes folding subagents into the main agent through skills and tools. Across several enterprise deployments, one agent with skills frequently delivered better quality and often lower cost and latency. That observation concerns those deployments and does not guarantee results for your operation.

Verdict by profile

• If your workflow shares an order, commercial terms and conversation, test one agent with skills before adding specialists.
• If you need independent research or must isolate a task that fills the context with irrelevant material, compare a version with subagents.
• If you cannot yet define a correct output or who handles exceptions, settle that before buying either architecture.

Our generalist vs specialized agents guide covers the broader criteria. The specific change here is turning each specialist’s procedure into material the same agent consults when needed.

Skills, tools and subagents in operations terms

A skill captures “how we work here”: steps, rules and judgment for a task. It can include templates, files and scripts. A tool retrieves data or performs an action. A subagent is another agent with its own context and instructions.

In Skills explained, published March 5, 2026, Anthropic describes progressive disclosure: brief descriptions appear first; full instructions and resources load when needed. This organizes procedures without pasting every one into a huge prompt. The October guide also applies deferred loading to tools through tool search.

At our fictional distributor, the request is: “Prepare a quote with availability and delivery information; leave it for review.”

Component	Proposed example
Quoting skill	Check the sales unit, apply the current commercial policy and flag missing data.
Delivery skill	Require warehouse confirmation before promising a date.
Tools	Retrieve the catalog, stock and authorized terms; save the draft in the system of record.
Possible subagent	Research alternative suppliers across lengthy documents and return options with sources.

The agent keeps the request and order decisions while consulting the procedures. If a confirmed date is missing, the draft says so. Written instructions do not replace the system control that prevents sending a quote without approval.

What each handoff costs

When one agent passes work to another, someone must decide which history, data and constraints travel with it. A summary might omit that the customer changed the purchase unit. Someone then has to spot the inconsistency and repeat the lookups.

Anthropic’s commerce agents guide, dated September 2, 2026, describes handoffs as operations that lose some state and add tokens and latency. Skills retain modularity inside the agent already handling the conversation.

Separate sales and warehouse owners do not prove that their tasks need separate contexts. Anthropic’s October guide also warns against reproducing the organization chart as subagents. In this example, both teams can maintain their procedures and review their cases, with one person accountable for the complete result.

When splitting does help

When to use multi-agent systems (and when not to), dated January 23, 2026, identifies three reasons: context pollution, tasks that can run in parallel, and specialization that improves tool selection or focus.

For the distributor, investigating several markets can become independent searches. If research adds documents that make the later quote worse, test isolating it and returning only necessary findings. If the agent confuses similar tools, compare clearer descriptions first and separation afterward.

Anthropic uses multiple agents in its own research system. In How we built our multi-agent research system, dated June 13, 2025, it reported a 90.2% improvement over single-agent Opus 4 on its internal research evaluation, using Opus 4 as lead and Sonnet 4 subagents. It also reported roughly 15 times the tokens of a chat, versus about four times for an agent. These results describe that system and evaluation, not a forecast for quoting. The article identifies a poorer fit where tasks share extensive context or have many dependencies.

Other boundaries can justify separation. A subagent may need restricted tools; verify actual isolation using our secure agent permissions guide. The commerce guide also allows a full handoff to a domain agent that needs to work directly with the user. For example, a request entering a dedicated financial services experience may need continuity within that domain.

Comparison at the same scope

This table compares designs for the same workflow. Commercial and hosting terms depend on the implementation; ask the provider for evidence.

Criterion	One agent with skills	Several agents
Verified entry price	There is no universal architecture fee. Request a quote for the defined workflow.	No universal fee here either. Require a breakdown for coordination and subagents.
Where processing and data live	Confirm hosting, region and retention for each integration.	Also confirm which data is copied between agents.
MCP status	The architecture does not determine this; verify whether the implementation uses it.	The agent count does not confirm MCP support either.
Authentication	Request the identity and scopes of each tool.	Also request identities and credentials for each agent.
Respects user permissions	Require a test that rejects access beyond the authorized scope.	Require it at each boundary; separate names do not demonstrate isolation.
Integrations	Inventory those needed for the complete task.	Also inspect duplicated connections and handoffs.
Limits	Set workflow budgets, time limits and retries.	Apply limits to the whole system as well as individual agents.
LATAM support	Confirm language, hours, billing and accountable owner in the contract.	Also confirm who handles a failure between agents.

Definitions follow the official guides linked above. We are not evaluating a product or assigning a particular region, plan or integration to an architecture.

Decision criteria: prepare an evaluation in an afternoon

Anthropic suggests starting with as few as ten examples and folding subagents in one at a time while comparing results. This is our operations adaptation; preparing the comparison in an afternoon does not mean finishing a migration that day.

1. Choose a verifiable output: a draft quote with authorized pricing, availability and a confirmed date or an explicit pending date. To narrow the task, use what to automate first.
2. Prepare about ten anonymized real cases. Include a normal order, insufficient stock, an ambiguous unit, expired commercial terms, missing data and a request exceeding permissions. Write the expected output before running them.
3. Compare the current design with a version where just one subagent becomes a skill and tools. Keep data, model, limits and criteria constant, and repeat cases to detect variation. Use a test environment without external sends.
4. Record acceptance per case, omissions, elapsed time, retries, usage and human correction time. Track which skill loaded and which tool was chosen without copying sensitive data into logs.
5. Review cases that got worse even if the average improved. Keep the subagent if the integrated version loses quality; document why before testing the next change.

Calculate cost per accepted task = total sample cost / completed and accepted tasks. The numerator also includes failed attempts and corrections. With no accepted tasks, the version fails the test. Our real cost of an AI agent guide develops the budget categories.

Keep the library small and distinct

In this example, “quoting” and “delivery” have different triggers. Two skills called “sales” and “commercial support” may compete for the same order. Write when to use each, the input it expects and who maintains its policy. Test selection on ambiguous cases before adding another.

Version procedures and examples together. When commercial terms change, identify which skill and cases need review. Our specialists in production guide covers work organization and review points.

What to ask the provider before paying

• Which specific single-agent failure requires each subagent? Request comparable cases and results.
• Can you inspect the input, output and information discarded at every handoff?
• Does the budget include retries, tools, observability, corrections and maintenance? How is it measured per accepted task?
• Where does the system block unauthorized actions, and can you test that rejection?
• Who maintains each procedure, and who is accountable for the complete result?
• Can a specialist be folded in or removed without rebuilding the workflow? Request a demonstration in a test environment.

When to choose neither

If stock has no reliable source, commercial policy contradicts itself or nobody can validate the output, either option will automate uncertainty. Resolve that dependency first. A data lookup, a rule and a manual draft may be enough to stabilize the process before the pilot.

Next step

Bring a diagram of the current workflow and two anonymized examples, one that works and one that fails, to a diagnostic conversation with Kiia. That gives us a basis to define what to compare and which evidence to request before changing the architecture.

All linked official sources were consulted on October 26, 2026. Their publication dates were visible during review. The examples and operating method are illustrative Kiia proposals, not benchmarks or promises of savings.