Skip to content

Placement rules: skill, subagent, or hand-off

Two decisions from Anthropic’s guide to commerce agents, drawn out. Figure 1 is where a capability runs. Figure 2 is where its instructions live.

Default to skills. A subagent or a hand-off earns its place only in the two exception cases. The deciding question is who owns the conversation.

One agent loads the instructions it needs.

User MAIN AGENT · ONE LOOP cart prefs history shared session state stays in view + customer-care skill instructions load into the same context
  • Use when: the work needs the session’s shared state: the cart or staged changes, preferences, history. Commerce sessions are tightly coupled across intents and turns, so this is most of them.
  • Cost: one model turn to load the skill. No handoff, so nothing is lost.
  • Example: a returns flow that needs order history, the current cart, and the catalog at once.

Messy, self-contained work in its own context window.

User MAIN AGENT keeps the conversation narrow brief SUBAGENT its own context window searches · reads · runs code dead ends stay in here compact answer
  • Use when: the orchestrator can call it as a tool for a narrow task that benefits from its own context window, and only a short result needs to come back.
  • Cost: each handoff loses state, can cost several times the tokens, and adds seconds of latency (the handoff tax). In the reference build it reads data but never adds to the IDs the agent may write to.
  • Example: deep research, or the merchant analysis subagent.

A purpose-built agent takes over the conversation.

MAIN AGENT steps back hands off the task User DOMAIN AGENT its own loop and compliance surface talks to the user until the task is done
  • Use when: the domain already runs a dedicated agent with its own compliance surface. That agent becomes the user’s counterpart.
  • Cost: one transfer of ownership. After that, the domain agent works through its own loop.
  • Example: a pharmacy or financial-services experience.
  1. Does a purpose-built agent with its own compliance surface already run this domain? (Pharmacy, financial services, anything with separate regulatory ownership.) Yes → Hand-off.
  2. Is it narrow and self-contained, with only a compact answer needed back? (Lots of searching, reading, and dead ends you don’t want in the main context.) Yes → Subagent as a tool.
  3. Otherwise, anything that touches the cart, preferences, or history mid-conversation → Skill (default).
Skill Subagent as a tool Hand-off
Talks to the user Main agent Main agent. The subagent never does. The domain agent, from the hand-off on
Context The main window, with the full history Its own window. Only the answer returns. Its own loop
Overhead One model turn to load Lossy handoff, several times the tokens, seconds of latency One transfer of ownership
Best for The long tail of domains in one session Deep research and analysis Domains with their own compliance surface
User INTENT ROUTER ✕ ✕ ✕ ✕ search subagent cart subagent returns subagent promos subagent
  • Every hop (✕) drops state the orchestrator was holding, which lowers the quality of the answer.
  • Each handoff can cost several times the tokens and adds seconds of latency.
  • Domains don’t separate cleanly. Returns needs order history, the cart, and the catalog, so access gets duplicated or the task gets handed off midway.
  • Splitting by business unit to ease team ownership has the same problem. Give each team ownership of its skills and tools instead.

In the authors’ comparisons across several enterprise deployments, one agent with skills beat both the one-prompt-for-everything design and the subagent design on quality, often at lower cost and latency per task. As models handle more context, skills, and tools, these limits loosen.

Figure 2: System prompt or skill, decided by frequency

Section titled “Figure 2: System prompt or skill, decided by frequency”

Loading a skill costs a model turn. So anything most turns need goes in the system prompt, and the long tail goes in skills. Two rules sit on top of that: critical rules always go in the prompt, and a predictable skill can be loaded before the turn starts.

Always in the system prompt, at any frequency

Section titled “Always in the system prompt, at any frequency”

Critical instructions skip the frequency test: safety rules, legal rules, brand constraints, and key user facts (such as allergies).

The frequency line: about a third of traffic

Section titled “The frequency line: about a third of traffic”
Skills: the long tail System prompt: what most turns need
Threshold Below about ⅓ of traffic. Loaded on demand. ⅓ of traffic or more, whether expected before launch or seen in production.
Shopping agent search-discovery, purchase-research, planning-goals, customer-care, memory-personalization Grounding, cart and checkout semantics, presentation rules, product search (nearly every session)
Merchant agent One per domain: performance-insights, catalog-listings, inventory-operations, pricing-promotions, marketing-campaigns

Predictable from a signal? The harness injects it

Section titled “Predictable from a signal? The harness injects it”

Load the skill before the first model call and skip the extra turn. The signal is something you already have, such as the page the user came from:

  • A shopper opened the assistant from a product page.
  • A merchant opened it from a campaign dashboard.
  1. Is it critical? Safety, legal, brand, key user facts → System prompt.
  2. Does about a third of traffic or more need it? → System prompt.
  3. Can a signal you already have predict it? → Harness preloads it.
  4. Everything else → Skill.

System prompt: sent on every request, cached

  • Lives in the global cache segment with the tool definitions. Keep it byte-identical across turns and sessions.
  • Cached reads cost a tenth of fresh input tokens.
  • Every line rides along on all traffic, so keep it for what most turns use.

Skill: one model turn to load

  • Load it as a tool result. Don’t append it to the system prompt.
  • The body then joins the conversation prefix and is cached with it.
  • Roll the cache breakpoint forward each turn so later rounds read it from cache.

The one-third line is a starting point. Where it falls depends on your traffic mix and on what your evals show the agent doing, so move instructions across it when production data says to.

  • A guide to the anatomy of effective commerce agents, Anthropic (Ali Shazal, Matthew Koen), Sept 2, 2026. Sections “Skills, not subagents” and “System prompt or skill: decide by frequency”.
  • Reference implementation: github.com/anthropics/commerce-agents.