It is easy to build more than a job needs: a team of agents where one agent with a skill would do. This guide helps you choose, using the decision questions from Anthropic’s guide Building effective AI agents: architecture patterns and implementation frameworks. The patterns are not specific to one product, so they apply to Claude Code, GitHub Copilot and OpenAI Codex alike.
Start with the simplest thing that works
Anthropic’s advice is to begin with a single-purpose agent that does one thing well, and grow only when it has to. Simple systems cost less to run, are easier to debug, and give you clear measures of whether they work. If you can not say why a second agent is needed, it is not.
Four questions
1. How much control do you need? If you must explain exactly why the system did something, for example for compliance or money, use a single agent or a fixed sequence of steps: their behaviour is easier to trace. With moderate control needs, such as support or content work, a lead agent that delegates to specialists can work. Where the work is open-ended, such as research or brainstorming, agents that collaborate become worth their unpredictability.
2. How complex is the work? One domain, such as answering questions or processing returns: a single agent. Several domains but a predictable path, such as onboarding or standard analysis: a sequential or parallel workflow. Complex and open-ended, such as research or troubleshooting: several agents.
3. What can you spend? Several agents use many times more tokens than one. Anthropic reports roughly 10 to 15 times, which is its own measurement, so check your own. Work out the cost at your expected volume before you commit. If you need results soon, start with one agent and plan how it could grow.
4. Do you need deep expertise? One domain with a well-known process: one agent with specialised skills. Several distinct domains that have to work together: several agents, each with its own skills.
The patterns, and what to use in Skill Builder
| Pattern | Use it when | In Skill Builder |
|---|---|---|
| Single agent with skills | Open-ended work in one domain | One agent, with the skills it needs linked or referenced |
| Sequential workflow | The work splits into fixed steps, and each step is easier on its own (draft, review, publish) | A skill or agent per stage, handing off in order |
| Parallel workflow | Independent checks or viewpoints can run at the same time, such as a safety screen beside the main task, or several reviews that are compared | Several read-only agents that a lead agent calls |
| Supervisor with specialists | One lead should split the work, send it to specialists and bring it together | A lead agent that references specialist agents |
| Evaluator–optimizer | One agent makes something and another judges it against clear criteria, round after round | The built-in Evaluator–optimizer workflow with a Completion goal |
Each has a case to avoid. Sequential steps are wrong for a task a single agent handles easily, or one that needs backtracking. Parallel work is wrong when agents depend on each other’s output or change shared state. An evaluator is wrong when the criteria are subjective, the first attempt is already good enough, or the answer is needed immediately. Anthropic notes that two to four rounds are typical.
Add skills before adding agents
Before moving to several agents, see whether one agent with a skill meets the need. A skill adds expertise and a repeatable workflow without a second agent to coordinate, and skills can be shared by several agents. Read Agent skills vs agents for the difference.
What each tool reads
Skills are read by Claude Code, Copilot and Codex. Agent files go in each tool’s own folder, and how each one delegates work differs, so check the tool’s own documentation before you design around it. Export skills and agents for Claude Code, Copilot and Codex lists the folders.
Grow in stages
Anthropic’s example is an online shop. It starts with one agent answering customer questions, then adds a routing step that separates order status, product questions and complaints. Next come specialised agents for each, then coordination between agents for inventory, payment and shipping, and finally agents that check quality. Each stage was added because the previous one worked and measurement showed a need. Your setup should evolve with your needs.
Keep it checkable
Whatever you pick, decide how you will know it worked:
- Give any repeating work a Completion goal: the goal, the check each round, a limit, and what happens next.
- Let a read-only agent judge another agent’s work, so it can not fix what it should judge.
- Write tests for each skill and agent. See How to test skills and agents.
- Save a version before you restructure, so you can compare and go back.
Common mistakes
- A team of agents for a job one agent could do.
- Agents that talk to each other with no one to settle a disagreement, or with no limit on bouncing a task back and forth.
- An evaluator with no criteria to judge by.
- Every agent preloading every skill, which fills the context of each one.
- Adding agents without measuring whether the first one was the problem.