36 Agents, One Pipeline

Why 36?

The number is not arbitrary. It maps to the real categories of work a software team does: write code, test it, review it, secure it, document it, refactor it, ship it, monitor it, and manage the process. One agent per specialization, all sharing the same context layer, all orchestratable by the PM engine.

The question is not "why do you need 36 agents?" — it is "why would you use fewer?"

Language Agents (10)

Each language agent is purpose-built for its ecosystem. It knows the idioms, the testing frameworks, the package managers, and the linting tools for its language.

Quality Agents (5)

These agents do not write new features — they verify, harden, and improve existing code.

• test-agent — Scans for untested code paths, generates comprehensive test suites, runs them, fixes failures. Reports coverage.
• review-agent — High signal-to-noise code review. Surfaces bugs, security issues, and logic errors. Never comments on style.
• security-agent — OWASP-aligned audit: injection, auth, secrets, input validation, dependency CVEs, supply chain risks.
• refactor-agent — Identifies code smells, plans safe refactoring sequences, executes with test verification after each change.
• debug-agent — Hypothesis-driven debugging. Traces execution, identifies root causes, implements minimal fixes, adds regression tests.

Infrastructure Agents (6)

• database-agent — Schema design, Alembic migrations, query optimization, seed data generation.
• devops-agent — GitHub Actions workflows, Dockerfiles, deployment pipelines, monitoring configs.
• cloud-agent — Terraform modules, Kubernetes manifests, AWS/Azure/GCP architecture.
• monitoring-agent — Health checks, structured logging, metrics, SLIs/SLOs, distributed tracing.
• performance-agent — CPU/memory/IO profiling, bottleneck identification, benchmark implementation.
• migration-agent — Alembic lifecycle management, zero-downtime migration planning, upgrade/downgrade testing.

Management Agents (7)

• pm-agent — Central dispatcher. Receives intent, decomposes tasks, coordinates all other agents.
• planning-agent — Scope definition, task decomposition, TODO.md generation, implementation plans.
• brainstorm-agent — Divergent idea generation. Explores 3–5 approaches, builds trade-off matrices.
• research-agent — Technology research, library comparison, prior art investigation. Feeds planning and brainstorm workflows.
• git-manager-agent — Merge conflict resolution, PR readiness checks, history cleanup.
• release-agent — Version bumping, changelog generation, git tagging, GitHub Release publishing.
• agent-generation-agent — Scaffolds new agents from a plain-language spec. Meta.

Documentation & Design Agents (3)

• docs-agent — Maintains DEV-LOG, CHANGELOG, TOC, READMEs, API docs, docstrings, architecture docs. Detects documentation drift.
• ui-ux-agent — Accessible interfaces, WCAG compliance, responsive layouts, design-to-code translation.
• design-systems-agent — Component libraries, design tokens, Storybook documentation, visual regression tests.

Utility Agents (5)

• dependency-agent — CVE scanning, license compliance, outdated package detection, supply chain risk analysis.
• scripts-agent — Cross-platform shell, Python, and PowerShell scripts with shared lib integration.
• project-audit — Holistic codebase analysis: code quality, security, tests, docs, architecture.
• data-science-agent — EDA, visualizations, data pipelines, statistical models (pandas, polars, scikit-learn).
• ci-cd — CI/CD pipeline configuration, validation, GitHub Actions, Docker deployment.

Composing Pipelines

Individual agents are useful. Pipelines are where CHAOS becomes something else entirely.

# Pre-commit gate (parallel test + security, then review)
cw run pm "pre-commit pipeline"

# Full feature lifecycle
cw run pm "feature complete: src/payments/ — refactor, test, document, review"

# Project health check
cw run pm "health check: test + security + docs"

# Release
cw run pm "release: bump version, changelog, tag, publish"

The PM engine handles dependency ordering, parallel dispatch, and result aggregation. You describe the goal. CHAOS runs the agents.

What This Means for a Solo Developer

With 36 agents sharing context and a PM engine coordinating them, a solo developer can run workflows that would normally require a team: implement a feature, generate its tests, audit for security issues, update the docs, review the diff, and open a PR — all in a single cw run pm command.

That is the point.