Skills, Sub-Agents, and Agent Swarms

Start here (if this is your first time coding)

If you are new to coding, terms like skills, sub-agents, and swarms can sound complicated. In practice, they are just ways to organize help from AI tools so your project stays understandable.

The key idea from Session 4 was simple: do not start by asking for a lot of code. Start by planning, decide what you want, and then execute in small steps you can check.

What each term means (in plain language)

1. Skills

A skill is a reusable instruction card. You write it once, then reuse it many times for similar tasks.

Example: Instead of retyping your full literature-review prompt every week, create one skill and call it when needed.

2. Sub-agents

A sub-agent is a specialized helper that does one focused job. You can use a sub-agent to review consistency, check style, or verify that data choices match your plan.

Example: One sub-agent checks whether your empirical code still matches the variable definitions chosen in your data strategy.

3. Swarms (multi-agent orchestration)

A swarm means multiple agents working together with different roles. This is useful for larger projects, but usually unnecessary at the beginning.

Example: One agent drafts literature notes, one prepares data, one performs quality checks, and one summarizes changes.

What beginners should do first

Pattern	Best used for	Typical risk if overused
Skills	Repeated tasks (great first step)	Can become too rigid if never updated
Sub-agents	Focused checks and reviews	Too many at once creates overhead
Swarms / multi-agent orchestration	Bigger projects with clear role separation	Confusion if roles are not clearly defined

For EC1B1 beginners, the recommended order is:

1. Start with one skill.
2. Add one sub-agent for consistency checks.
3. Only explore swarms when your project structure is already clear.

The AGENTS.md standard (important)

What AGENTS.md is

AGENTS.md is a project-level instruction file that explains your goals, constraints, and workflow rules to the AI tool. Think of it as your project constitution.

It helps prevent drift when you run multiple tasks over time.

Why this matters for first-time coders

When you are learning to code, it is easy to lose track of assumptions. For example, you may choose quarterly data in one part of the project and accidentally use annual definitions somewhere else.

A clear AGENTS.md file reduces this problem by giving every agent the same reference point.

Minimal AGENTS.md template

# AGENTS.md

## Project Goal
State the research question in 2-3 lines.

## Scope
What is in scope this week? What is explicitly out of scope?

## Data Definitions
List main variables, frequency, sample period, and preferred sources.

## Workflow Rules
1. Plan before execution.
2. Implement in small steps.
3. Explain changes in plain language.
4. Run consistency checks before final output.

## Quality Checks
- Confirm variable definitions are consistent across files.
- Confirm code results match written interpretation.
- Flag uncertainty instead of guessing.

## Output Format
Specify expected deliverables (files, summary, tables, plots).

Cross-tool note

Many tools support an AGENTS.md-style anchor file. For Claude Code specifically, the equivalent project memory pattern is often maintained in CLAUDE.md.

Recommended workflow for your coursework

1. Write a short AGENTS.md before coding.
2. Ask for a plan first.
3. Approve or edit the plan.
4. Execute one chunk at a time.
5. Use one sub-agent to run a consistency check.
6. Review and verify manually before submission.

Primary references

Skills

• Cursor skills: cursor.com/docs/context/skills
• Claude Code skills: code.claude.com/docs/en/skills

Sub-agents

• Cursor subagents: cursor.com/docs/context/subagents
• Claude Code sub-agents: code.claude.com/docs/en/sub-agents

Project instruction files

• Cursor rules/project instructions: cursor.com/docs/context/rules
• Claude project memory (CLAUDE.md): code.claude.com/docs/en/claude-md

Swarms and orchestration (advanced)

• OpenAI Agents SDK (Python): openai-agents-python docs
• Multi-agent patterns (OpenAI Agents SDK): multi_agent
• OpenAI Swarm (experimental lightweight framework): github.com/openai/swarm
• Microsoft AutoGen: microsoft.github.io/autogen/stable
• CrewAI docs: docs.crewai.com

EC1B1 safety checklist

Before you accept final output, confirm these four points:

1. Your research question is unchanged.
2. Your data definitions are consistent across files.
3. Your code still matches your empirical design.
4. Your written interpretation matches the actual output.

If one point fails, pause and run a consistency pass before continuing.