A comprehensive collection of generic agent templates and patterns for Devin AI multi-agent systems. These templates provide a foundation for building specialized sub-agents while keeping project-specific knowledge private.
This repository contains generic agent templates, patterns, and documentation that can be customized for any project. The templates are designed to be copied to your local Devin configuration and then adapted to your specific needs.
The system uses a hierarchical delegation pattern where a coordinator agent orchestrates specialized sub-agents for specific tasks. The ecosystem is self-improving via a sensor/actuator loop.
graph TD
USER[User Request] --> GC{global_coordinator}
GC -->|Python| PC[python_coordinator]
GC -->|Other| COORD[coordinator<br/>generic fallback]
GC -->|Planning| PL[planner]
PC --> IMPL[Implementation Specialists]
COORD --> IMPL
IMPL --> QUAL[Quality & Safety]
QUAL --> GW[git-workflow]
GW --> REVIEW[Human Review]
USER -.->|gap detected| SR[subagent-recommender<br/>sensor]
SR -.->|proposes| SC[subagent-curator<br/>actuator]
SC -.->|creates/edits| ECOSYSTEM[Agent Ecosystem]
style GC fill:#ff6b6b,stroke:#c92a2a,stroke-width:2px,color:#fff
style SR fill:#e599f7,stroke:#862e9c,stroke-width:2px,color:#fff
style SC fill:#e599f7,stroke:#862e9c,stroke-width:2px,color:#fff
style REVIEW fill:#fcc419,stroke:#e67700,stroke-width:2px,color:#000
See agent-architecture.md for detailed diagrams (delegation flow, verification pipeline, improvement loop) and full agent/skill inventory.
- coordinator-template: Main orchestrator for task delegation and synthesis
- python-reviewer-template: Code quality and review specialist
- api-specialist-template: API design and implementation
- testing-guardian-template: Test quality and validation
- streamlit-expert-template: UI architecture and Streamlit applications
- redis-engineer-template: Caching strategies and Redis integration
- ollama-specialist-template: LLM integration and Ollama management
- devops-docker-template: Container orchestration and deployment
- security-auditor-template: Security scanning and vulnerability assessment
- git-workflow-template: Branch management and commit workflows
.
├── templates/ # Generic agent templates
│ ├── coordinator-template.md
│ ├── python-reviewer-template.md
│ ├── streamlit-expert-template.md
│ ├── redis-engineer-template.md
│ ├── ollama-specialist-template.md
│ ├── security-auditor-template.md
│ ├── git-workflow-template.md
│ └── etc.
├── patterns/ # Generic patterns and workflows
│ ├── coordinator-optimization-patterns.md
│ ├── coordinator-quick-reference.md
│ ├── task-template-patterns.md
│ ├── task-templates-quick-reference.md
│ ├── skills-integration-patterns.md
│ ├── skills-quick-reference.md
│ ├── delegation-patterns.md
│ ├── streamlit-performance.md
│ └── redis-patterns.md
├── plugins/ # Devin plugin (installable unit)
│ └── devin-agents/
│ ├── agents/ # 22 subagent profiles
│ ├── skills/ # 29 skills
│ ├── rules/ # 2 triggered rules
│ └── AGENTS.md # always-on rule
├── scripts/ # Utility scripts
│ ├── install-agents.sh
│ └── validate-agent.sh
├── agent-architecture.md # Architecture documentation
├── CUSTOMIZATION.md # How to customize templates
├── PLAN.md # Planning template
└── README.md # This file
This repo also ships a Devin plugin at plugins/devin-agents/ that bundles
the full multi-agent team into a single installable unit. The plugin merges every
unique agent and skill from the global Devin config (~/.config/devin/) and
project repos into one deduplicated, generic package.
- 22 custom subagent profiles (
agents/<name>/AGENT.md) — coordinators, implementation specialists, reviewers, workflow agents, and a meta-agent (subagent-curator) that maintains the agent ecosystem itself. - 29 skills (
skills/<name>/SKILL.md) — invokable skills including thesubagent-recommender(detects coverage gaps and proposes new sub-agents),subagent-curator(reviews, edits, creates, and audits profiles), and 6 process skills (grilling,tdd,diagnosing-bugs,code-review,codebase-design,handoff) adapted from engineering best practices. - Always-on rule (
AGENTS.md) — installs the coordinator-first workflow, the auto-recommend guidance, and the continuous improvement loop in every session. - Triggered rules (
rules/) —subagent-recommender.mdprompts the agent to invoke the recommender skill when it sees repeated, unscoped, or cross-cutting work;continuous-improvement.mdties the recommender and curator together into a self-improving ecosystem.
plugins/devin-agents/
├── .devin-plugin/
│ └── plugin.json # manifest
├── AGENTS.md # always-on rule (coordinator workflow + improvement loop)
├── rules/
│ ├── subagent-recommender.md # triggered rule: detect coverage gaps
│ └── continuous-improvement.md # triggered rule: ecosystem self-improvement
├── agents/ # 22 subagent profiles
│ ├── global_coordinator/AGENT.md
│ ├── coordinator/AGENT.md
│ ├── subagent-curator/AGENT.md # meta-agent: reviews/edits/creates profiles
│ └── … (19 more)
└── skills/ # 29 skills
├── subagent-recommender/SKILL.md # detect gaps, propose new sub-agents
├── subagent-curator/SKILL.md # review/edit/create/audit profiles
├── grilling/SKILL.md # pre-implementation interview
├── tdd/SKILL.md # red-green-refactor discipline
├── diagnosing-bugs/SKILL.md # 6-phase debugging
├── code-review/SKILL.md # two-axis review (standards + spec)
├── codebase-design/SKILL.md # deep module design vocabulary
├── handoff/SKILL.md # session continuity
├── coordinator/SKILL.md
└── … (20 more)
# From the repo root
devin plugins install ./plugins/devin-agents
# Or from any location
devin plugins install /path/to/devin-agents/plugins/devin-agentsLocal installs are linked, so edits to the plugin files apply on the next session
— no update needed. Verify with:
devin plugins list
devin plugins info devin-agentsSkills become available as /devin-agents:<skill> slash commands. Subagent
profiles are available to run_subagent by name.
Process skills run inline in the current conversation — the skill's instructions are injected and the agent follows them directly:
/devin-agents:grilling I want to add a caching layer to the API
/devin-agents:tdd implement a cache wrapper with TTL support
/devin-agents:diagnosing-bugs the API returns 500 on large payloads
/devin-agents:code-review main
/devin-agents:codebase-design
/devin-agents:handoff next session should focus on integration tests
Subagent-tied skills spawn a specialist subagent with its own context window, tools, and model. The parent agent waits for the result and summarizes it:
/devin-agents:python-reviewer src/services/
/devin-agents:security-auditor
/devin-agents:architecture-reviewer
/devin-agents:testing-guardian tests/
You can also ask the agent to use a skill in natural language:
"I want to review this code" → the agent reaches for /devin-agents:code-review
"Debug this issue" → the agent reaches for /devin-agents:diagnosing-bugs
22 subagent profiles (invoke by name via run_subagent):
| Category | Profile | Focus |
|---|---|---|
| Coordinators | global_coordinator |
Detects language/stack, delegates |
coordinator |
Generic orchestrator | |
python_coordinator |
Python-specific orchestrator | |
planner |
Spec/PLAN.md production | |
| Implementation | python-developer |
FastAPI/Flask/Django backend |
api-specialist |
REST endpoints, OpenAPI | |
streamlit-expert |
Streamlit UI, caching, reruns | |
redis-engineer |
Redis caching, resilience | |
ollama-specialist |
Ollama LLM, streaming | |
devops-docker |
Docker Compose, deployment | |
| Quality | python-reviewer |
Python code review |
swe-check |
Non-Python bug detection | |
security-auditor |
Vulnerabilities, secrets | |
testing-guardian |
Test coverage, quality | |
qa-ci-agent |
CI/CD gates, lint, typecheck | |
architecture-reviewer |
Module boundaries, structure | |
best-practices-reviewer |
Cross-language code quality | |
feature-verifier |
Verify features match spec | |
| Workflow | git-workflow |
Branches, commits, merges |
documentation-agent |
Docs, README, migration guides | |
playwright-testing |
E2E test creation | |
| Meta | subagent-curator |
Reviews/edits/creates profiles |
29 skills (invoke as /devin-agents:<skill>):
| Category | Skill | Description |
|---|---|---|
| Coordinator | /devin-agents:coordinator |
Pure orchestration |
| Implementation | /devin-agents:python-developer |
Python backend |
/devin-agents:api-specialist |
API design | |
/devin-agents:streamlit-expert |
Streamlit UI | |
/devin-agents:redis-engineer |
Redis caching | |
/devin-agents:ollama-specialist |
Ollama integration | |
/devin-agents:devops-docker |
Docker/DevOps | |
| Quality | /devin-agents:python-reviewer |
Python review |
/devin-agents:architecture-reviewer |
Architecture review | |
/devin-agents:security-auditor |
Security audit | |
/devin-agents:testing-guardian |
Test quality | |
/devin-agents:qa-ci-agent |
CI/CD gates | |
/devin-agents:swe-check |
Non-Python bugs | |
/devin-agents:best-practices-reviewer |
Cross-language quality | |
/devin-agents:feature-verifier |
Verify features match spec | |
| Workflow | /devin-agents:git-workflow |
Git operations |
/devin-agents:documentation-agent |
Documentation | |
/devin-agents:playwright-testing |
Playwright tests | |
| Audits | /devin-agents:ollama-testing |
Ollama safety audit |
/devin-agents:redis-resilience |
Redis resilience audit | |
/devin-agents:quick-review |
Quick pre-commit review | |
| Meta | /devin-agents:subagent-recommender |
Detect gaps, propose agents |
/devin-agents:subagent-curator |
Review/edit/create/audit profiles | |
| Process | /devin-agents:grilling |
Pre-implementation interview |
/devin-agents:tdd |
Red-green-refactor discipline | |
/devin-agents:diagnosing-bugs |
6-phase debugging | |
/devin-agents:code-review |
Two-axis review (standards + spec) | |
/devin-agents:codebase-design |
Deep module design vocabulary | |
/devin-agents:handoff |
Session continuity |
The plugin's subagent-recommender skill + rule makes Devin proactively suggest
new sub-agent profiles when it detects:
- The same task type handled inline 3+ times with no matching specialist
- Cross-cutting work that would benefit from an isolated context
- A coverage gap no existing profile fills
- An explicit user request for a new specialist
The skill drafts a complete AGENT.md (name, model, tools, system prompt) and
presents it for approval. On approval, the subagent-curator agent creates and
validates the profile.
The plugin includes a self-improving ecosystem via two complementary components:
subagent-recommender(sensor) — detects coverage gaps and proposes new profilessubagent-curator(actuator) — reviews, edits, creates, and audits profiles for consistency, genericity, quality, and coverage
The continuous-improvement.md rule ties them together with automatic triggers:
- 5+ subagent calls in a session → suggest a lightweight audit
- 3+ corrections to the same agent → its profile needs refinement
- Stale references or genericity drift → curator fixes them
- User request → run a full audit or improvement cycle
Usage:
/devin-agents:subagent-curator audit # full ecosystem audit
/devin-agents:subagent-curator improve # top improvements cycle
/devin-agents:subagent-curator edit <name> # edit a specific profile
/devin-agents:subagent-curator create <name> # create a new profile-
Clone this repository:
git clone https://github.com/moltra/devin-agents.git cd devin-agents -
Install templates to your local Devin config:
bash scripts/install-agents.sh
-
Customize for your project:
- Edit agent configurations in
~/.config/devin/agents/ - Add project-specific patterns and knowledge
- Adjust permissions and tool access
- Update model assignments if needed
- Edit agent configurations in
-
Restart Devin to load the new configurations
If you prefer manual installation:
# Copy agent templates
cp -r templates/* ~/.config/devin/agents/
# Install the plugin (recommended — includes all agents + skills)
devin plugins install ./plugins/devin-agents
# Make scripts executable
chmod +x scripts/*.shAfter installing the templates, you should customize them for your project:
-
Add Project Knowledge:
- Add project-specific file paths
- Include business logic patterns
- Document internal architecture
- Add configuration details
-
Adjust Permissions:
- Add project-specific tool permissions
- Configure file access patterns
- Set up Docker/container permissions
- Configure API access
-
Update Models:
- Choose appropriate models for your use case
- Adjust model parameters
- Configure model-specific behaviors
See CUSTOMIZATION.md for detailed customization guidelines.
coordinator → domain-specialist → python-reviewer → testing-guardian → security-auditor → coordinator
coordinator → api-specialist → python-reviewer → testing-guardian → security-auditor → git-workflow
coordinator → devops-docker → security-auditor → coordinator
coordinator → subagent_explore → domain-specialist → python-reviewer → coordinator
- Single Responsibility: Each agent has a clear, focused domain
- Hierarchical Delegation: Coordinator breaks down tasks, specialists execute
- Parallel Execution: Independent subtasks run concurrently when possible
- Cross-Validation: Code reviewed by multiple specialists
- Safe Operations: Git and destructive operations have limited permissions
- Privacy First: Project-specific knowledge stays in your local configuration
- CUSTOMIZATION.md: How to customize templates for your project
- agent-architecture.md: Multi-agent architecture documentation
- patterns/coordinator-optimization-patterns.md: Coordinator behavioral patterns for efficiency
- patterns/coordinator-quick-reference.md: Quick reference for coordinator behavior
- patterns/task-template-patterns.md: Task construction templates
- patterns/task-templates-quick-reference.md: Quick reference for task templates
- patterns/skills-integration-patterns.md: Skills integration guide
- patterns/skills-quick-reference.md: Quick reference for available skills
- patterns/delegation-patterns.md: Structural delegation patterns
- patterns/streamlit-performance.md: Streamlit optimization patterns
- patterns/redis-patterns.md: Redis integration patterns
When contributing new templates or patterns:
- Keep them generic and reusable across projects
- Remove any project-specific information (paths, configs, business logic)
- Add clear documentation for customization
- Update relevant documentation files
- Test the template before submitting
This repository contains only generic templates and patterns.
What IS shared:
- Generic agent configurations
- Reusable patterns and workflows
- Best practices and guidelines
- Documentation and examples
What is NOT shared:
- Project-specific file paths
- Business logic details
- Proprietary algorithms
- Internal architecture diagrams
- Configuration values (API keys, secrets)
- Custom agent implementations
Keep your project-specific customizations in your local ~/.config/devin/agents/ directory - these should never be committed to this repository.
Found a bug, have a feature request, or want to suggest a new agent profile?
- Bug reports — describe what happened, what you expected, and how to reproduce
- Feature requests — describe the use case and what agent/skill would help
- Agent proposals — use the subagent-recommender skill, then share your proposal in an issue for inclusion in the plugin
- General feedback — all feedback welcome
Please include:
- Your Devin CLI version (
devin --version) - The plugin version (
devin plugins info devin-agents) - Steps to reproduce (for bugs)
This project is licensed under the MIT License - see the LICENSE file for details.
For issues or questions:
- Review CUSTOMIZATION.md for guidance
- Check the agent-architecture.md for architecture details
- Open an issue on GitHub for template-specific problems