Agent Decomposition: Skills vs Tools

Claude skills, tools, and subagents do different jobs. Learn where to put instructions, actions, and separate workers.

A slow agent does not necessarily need a shorter prompt. This illustrated teardown shows how to divide a workflow: load instructions only when needed, call tools for evidence and action, and delegate only when separate context earns its cost.

Follow an inventory assistant from policy to stock evidence, draft recommendation, and human approval. Then examine Anthropic's StockPilot workshop: a 402-line system prompt became 15 lines, while its daily sweep went from 102 tool calls to 3 scripts and from 488 seconds to about 100. Those are workshop-reported results from several changes together—not a benchmark reproduced here or a guaranteed speedup.

The test is the complete workflow: correctness, elapsed time, and usage. A faster wrong answer is not an improvement.

YOUR TURN Paste into Claude: "Split my workflow into instructions, actions, and separate workers. Justify each boundary." Check: can every action be tested? Does every separate worker need its own context?

CHAPTERS 0:00 Divide the work, not just the prompt 0:10 Skills, tools, and subagents 0:22 What belongs in the core? 0:36 402 lines to 15: where the instructions went 0:51 Skills load into context 1:05 Tools perform actions 1:19 When a separate worker helps 1:34 An inventory workflow 1:48 Human approval before spending 2:02 Batching: 102 calls to 3 scripts 2:16 What the reported speedup actually means 2:32 Measure the whole workflow 2:46 Your turn: justify the boundaries

SOURCES Anthropic StockPilot workshop: https://github.com/anthropics/cwc-workshops/blob/main/agent-decomposition/README.md Claude Code skills: https://code.claude.com/docs/en/skills Claude Code subagents: https://code.claude.com/docs/en/sub-agents

CREDITS Nik Bear Brown. Narrated by Liam, in for Bear, using a synthetic Kokoro voice. Original animated illustrations; the inventory scenario is a teaching example. Independent educational commentary, not an Anthropic endorsement.

What this video covers

• 0:00 — Divide the work, not just the prompt
• 0:10 — Skills, tools, and subagents
• 0:22 — What belongs in the core?
• 0:36 — 402 lines to 15: where the instructions went
• 0:51 — Skills load into context
• 1:05 — Tools perform actions
• 1:19 — When a separate worker helps
• 1:34 — An inventory workflow
• 1:48 — Human approval before spending
• 2:02 — Batching: 102 calls to 3 scripts
• 2:16 — What the reported speedup actually means
• 2:32 — Measure the whole workflow
• 2:46 — Your turn: justify the boundaries

Transcript

A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.

Divide the work, not just the prompt

[0:00] Hello, this is Liam in forbear. A slow agent does not necessarily need a shorter prompt. It needs a better division of work. Let's separate skills, tools, and sub aents. Three different

Skills, tools, and subagents

[0:11] Things. A skill supplies instructions. A tool performs an action. A sub aent works in a separate context. A skill can guide tool use, but that does not automatically make it a separate agent.

What belongs in the core?

[0:22] Imagine putting every procedure into the agents starting instructions, forecasting, checking inventory, writing a report. Most tasks need only some of that material. The first design question is what belongs in the core and what can wait until it is needed. Anthropic stock

402 lines to 15: where the instructions went

[0:37] Pilot workshop reports reducing its system prompt from 42 lines to 15. The other instructions did not disappear. Roughly 400 lines moved into skills loaded on demand. That is deferred context, not deleted knowledge. Here is

Skills load into context

[0:52] A skill being used. The agent loads the relevant instructions into its existing context. The page joins the current work. By default, this is not a fresh worker with a private memory. Separate execution has to be configured explicitly. Now the agent calls a tool.

Tools perform actions

[1:07] A request goes out. A result comes back. That tool might read inventory, run code or write a file. Tools are not automatically harmless or stateless. What they can change depends on the implementation and permissions. A sub

When a separate worker helps

[1:20] Agent is different. Give a worker a bounded task in its own context. It does the work there, then returns a result. This can keep intermediate detail out of the parents conversation. It can also add coordination time and additional model usage. Try an inventory assistant.

An inventory workflow

[1:35] A skill explains the reorder policy. A tool reads the stock record. The agent uses both to draft a recommendation. The instruction page tells it how to decide. The tool result supplies evidence. Neither replaces the other. Before

Human approval before spending

[1:49] Placing an order, put a human approval point between the draft and the action. Check the quantity, supplier, and evidence. Reading stock and spending money should not inherit the same permission merely because they sit inside the same workflow. The workshop

Batching: 102 calls to 3 scripts

[2:03] Also changed how data was processed. Its daily sweep went from 102 tool calls to three scripts. Those scripts worked over data directly. Moving instructions into skills was only one part of the redesign. Batching work was another. For

What the reported speedup actually means

[2:16] That daily sweep, the workshop reports 488 seconds before and about 100 seconds after. Several changes contributed. Treat this as a reported example, not a guaranteed five-fold speed up for your agent. We have not rerun that benchmark here. Test your redesign on the same

Measure the whole workflow

[2:33] Tasks. Check correctness, elapse time, and usage. If the result is faster but wrong, reject it. If a separate worker returns a huge transcript, you may just have moved the clutter. Measure the complete workflow, not the prompts line count. Your turn. Paste this into

Your turn: justify the boundaries

[2:47] Claude. Split my workflow into instructions, actions, and separate workers. Justify each boundary. Then check two things yourself. Can every action be tested? Does every separate worker need its own context? If not, simplify the design. Agent decomposition skills vs tools at Nick Bear Brown.