Guide
Claude skills vs agents vs subagents: which one you need
A skill, a subagent and an agent sit at three levels of handing work over. A skill teaches Claude how to do one task, and loads inside the chat you're in when the task matches; you still direct the work. A subagent in Claude Code is a separate worker with its own context window, instructions and tools, which Claude hands a matching task to and which returns only the result. An agent decides its own steps and tool use until a goal is met. They combine: an agent can use a skill for each item and a subagent for side work. Anthropic's advice is to choose the simplest that works, so most teams start with skills, and write a harness before letting an agent run.
Short answer
- Skill: how to do a task, in your chat.
- Subagent: a side worker with its own context.
- Agent: decides its own steps toward a goal.
- Start simple; write a harness before an agent runs.
Side by side
| Skill | Subagent | Agent | |
|---|---|---|---|
| Who directs | You, in the chat | Claude hands it one task | It decides its own steps |
| Context | Your conversation | Its own window; returns a summary | Its own, across many steps |
| Good for | A repeated task done the same way | Research or checks that would flood the chat | Many similar items toward one goal |
| What it needs first | A clear description and steps | A description and the tools it may use | A done test, limits, stop rules and approvals |
From Anthropic's engineering post, the Claude Code docs and the Claude Help Center, read 5 November 2026.
A worked example: an agent that uses a skill
Illustration: a made-up business. An online shop has 30 product pages missing care instructions. The team has a product-copy skill that writes them in its voice. Before letting an agent fill all 30, the owner writes the harness:
From the harness
Done test: each of the 30 pages has care instructions of 40 to 80 words, from the supplier's sheet. The agent may mark items done, never change the list.
Tools: read the supplier sheets folder; write drafts to the drafts folder only. No publishing.
Memory: done.md and progress.md, read at the start and written at the end of every session.
Steps: one page at a time: read progress, pick the next page, write it with the product-copy skill, check the word count and the sheet, record it.
Stop: all 30 pass; the same page fails three times; a sheet is missing.
Human gate: the owner publishes every page.
Pilot: 5 pages first, every log line read.
In the pilot, two drafts used a washing temperature that wasn't on the sheet. The owner added a rule to the skill (only what the sheet says) and re-ran the five before the other 25.
Choose
- A task you repeat and check yourself: a skill.
- A side task that would flood the chat in Claude Code: a subagent.
- Many items toward one goal, with output you can check: an agent with a harness, using skills for each item.
- Anything sent, paid, deleted or published waits for a person.
Next step: sharing the skills agents use
Agents and people should use the same copy of each skill. The cheapest way to share AI prompts with a small team compares the options. For the terms, see AI agents and agent harness.
Skills for this
- AI practiceHarness spec for an AI agentWrites a one-page harness for an AI agent before it runs: done test, tools and limits, progress notes, small steps, stop rules, human approvals, and logs.
- AI practiceTurn a prompt into a skillTurns a prompt that works into a skill the whole team can reuse, with a valid name, a when-to-use description, steps, an output template, an example, and checks.
- AI practiceAI automation opportunity scanScores a team's recurring tasks on frequency, time, rule clarity, and risk to find AI candidates, then picks the top 3 with a first experiment each.
- AI practiceEval set for an AI taskBuilds a 20-to-50-case eval set for an AI task: normal, edge, and adversarial cases, expected results or rubrics, pass criteria, and a re-run routine.
Questions
What's the difference between a Claude skill and an agent?
A skill is a set of instructions Claude loads when a task matches, inside the chat you're in; you still direct the work. An agent decides its own steps and which tools to use until a goal is met. An agent often uses skills: the skill says how to do each item, the agent decides which item is next.
What's a subagent in Claude Code?
A separate worker with its own context window, instructions and tool access. Claude Code hands it a task that matches its description, and it returns only the result, so a side task doesn't fill the main conversation.
Should a small team start with agents?
Usually not. Anthropic's advice is to find the simplest solution possible and add complexity only when needed. Start with skills a person runs and checks; give a task to an agent only when its output can be checked and its mistakes are cheap, with a written harness.
AI skill
Have your AI write the rules an AI agent runs by
Put this skill in AI Skill Pod and everyone on your team gets it in their Claude, Cursor or ChatGPT (Pro, Business, Enterprise): one invite link, free to start. Harness spec for an AI agent has your AI write the one-page harness before an agent runs: a done test, the tools and their limits, progress notes, one item per step, stop rules, human approvals, and a log the owner can read in five minutes.
14 days free · No card needed
Sources
- Anthropic: Building effective agents (19 December 2024) — workflows follow predefined code paths; agents "dynamically direct their own processes and tool usage"; find the simplest solution possible (read 2026-11-05)
- Claude Code docs: Create custom subagents — each subagent runs in its own context window with a custom system prompt, specific tool access and independent permissions; Claude delegates when a task matches its description (read 2026-11-05)
- Claude Help Center: What are skills? — Claude reviews available skills, loads relevant ones and applies their instructions (read 2026-11-05)