Skip to content

JahnelGroup/multi-agents

Repository files navigation

multi-agents

Shared multi-agent pipeline setup for Cursor. Provides standard agents, rules, and pipeline tooling that teams copy into their projects. Claude Code support is documented alongside Cursor conventions.

Prerequisites

  • Cursor with models enabled for the agents in the pipeline. Some models (e.g. gpt-5.1-codex-max) are hidden by default and must be toggled on in Cursor Settings > Models. See Models | Cursor Docs for the full list and visibility defaults.
  • Python 3.10+, Node.js 20+, and git for tutorials and the sandbox project. See TESTING.md for details.

Getting started

Can you explain what agents, rules, and pipelines are? If not, start with Foundation.

Can you set up a pipeline and build a feature end-to-end? If not, go to Practitioner.

Ready to design multi-agent systems and lead others? See Expert.

The .cursor-foundation/, .cursor-practitioner/, and .cursor-expert/ directories are the canonical source bundles in this repo. Use them for training and for copying into your project.

Tier directory layout

Each .cursor-<tier>/ directory follows this structure (not all tiers include every subdirectory):

Path Purpose
agents/*.md One file per agent defining its model, role, inputs, outputs, and behavioral instructions. Cursor discovers these as subagent_type targets.
rules/*.mdc Always-on or file-triggered behavioral guardrails injected into agent context automatically. YAML frontmatter sets description, alwaysApply, and optional globs.
skills/*/SKILL.md On-demand capabilities agents pull in when relevant (unlike rules, not auto-injected). Each skill is a subdirectory with a SKILL.md. Foundation has none.
pipeline/ Runtime tooling: schema.py (validates artifact JSON), check.py (stage-gate invariants), README.md (artifact format docs).
templates/ Scaffolds for new agents, rules, and artifacts. Copy and fill in. Foundation has none.
walkthrough/ Pre-built example artifacts showing a complete pipeline run. Read-only reference material that ships with the repo. Foundation has none.
tutorials/ Learner output directory (outputs/), solution keys (solutions/), and verify.py grader. Exercise instructions live in docs/.
AGENTS.md Agent registry: index table of all agents in the tier, pipeline execution order, and subagent type mappings.
README.md Tier landing page: competency question, learning objectives, glossary (Foundation), pipeline flow, and tutorial links.

Competency framework

Foundation Practitioner Expert
Core Question Can you understand and use AI effectively? Can you build and deploy AI features? Can you architect AI systems and lead others?
Analogy Understanding the rules and controls Being able to drive anywhere safely Teaching others to drive and designing better roads
Expectation Understand the concept and patterns Experiment with multi-agent frameworks Design and deploy multi-agent systems with monitoring
Portfolio 3 documented AI use cases 1 deployed AI use case 1 client architecture, 1 presentation, 1 mentorship
Assessment Conversation with Practitioner or Expert Technical demo + walk-through with Expert Peer review + mentorship vouching
Contains 3 agents 1 rule 6 exercises 8 agents 4 rules 2 skills full pipeline 11 exercises ~15 tiered agents routing rule cost tracking 8 exercises

Documentation references

This repo builds on official Cursor and Claude Code documentation:

Cursor Learn -- concept guides for working with agents:

Cursor Docs -- technical reference:

  • Custom Agents -- agent .md files, AGENTS.md, subagent_type
  • Rules -- .mdc rule files, frontmatter, activation
  • Agent Skills -- SKILL.md format, discovery, on-demand activation

See the Claude Code section below for equivalent concepts in that system.

Adoption

Copy .cursor-<tier>/ into your project as .cursor/. Most teams should start with Practitioner.

cp -r .cursor-practitioner/* your-project/.cursor/

Tutorials

Each tier includes a tutorials/ directory with exercises. Foundation exercises are quiz-based and conceptual. Practitioner exercises are hands-on and require delegating to subagents. Expert exercises involve tiered routing, cost analysis, and architecture design. See TESTING.md for the full test plan.

Upgrading tiers

  • Copy .cursor-practitioner/ (or expert) over your .cursor/
  • Keep jg- files as read-only upstream references
  • Put customizations in team- or my- prefixed files so they survive upgrades
  • Diff before overwriting if you've modified any jg- files

Naming convention

Prefix Meaning
jg-* Shared bundle — do not modify in your project
<team>-* Team or project conventions
Unmarked Individual developer additions

Claude Code

Pipeline concepts (artifacts, agent roles, stage gates) are IDE-agnostic. The wiring differs:

Cursor Claude Code
.cursor/rules/*.mdc CLAUDE.md at repo root
.cursor/agents/*.md Referenced docs in CLAUDE.md
.cursor/skills/*/SKILL.md .claude/commands/*.md
subagent_type dispatch Sequential prompting through stages

Walkthrough content and pipeline artifacts work in both environments.

Security and guardrails

  • Agents that can write code: worker, debugger
  • Agents that can run commands: tester, git
  • Agents that cannot: merge PRs, force push, skip hooks, push to main
  • readonly: true on planner, reviewer, subplanner
  • Always review agent-generated PRs before merging

.gitignore

.pipeline/ is runtime state — do not commit. Walkthrough artifacts in tier directories live under walkthrough/, not .pipeline/, so they ship with the repo.

Cost awareness

Each pipeline run invokes multiple AI models. See Expert tier for detailed cost analysis and tiered model strategies.

Migration from .cursor-jg

If your project previously used .cursor-jg/ references, replace all paths with .cursor/. The old path convention is retired.

Versioning

VERSION file at repo root (semver). Bumped on any change to agents, rules, skills, or pipeline.

Maintenance policy

Tier directories are canonical in this repo. Update the affected tier directories directly and keep shared files synchronized across tiers when applicable.

Sync checklist:

  • Agents shared across tiers: planner, worker, git (Foundation, Practitioner, Expert)
  • Agents in Practitioner + Expert: subplanner, tester, reviewer, debugger, benchmarker
  • Rules shared: planner-first (all), commit-conventions, issue-workflow, pr-review (Practitioner, Expert)
  • Skills shared: pipeline-artifact-io, benchmark-ops (Practitioner, Expert)
  • Pipeline: README.md, schema.py, check.py (Practitioner, Expert — Expert extends)