ai-agents-workflow
Builds and implements AI agents, automation workflows, and tool integrations — MCP servers (2025 async, OAuth, stateless spec), LangChain and CrewAI agents, multi-LLM orchestration, and no-code platfo
SKILL.mdplugins/loxihart-ai/skills/ai-agents-workflow/SKILL.md
GitHubEvals
- ✅ SHOULD TRIGGER — "Build a LangGraph agent with memory and tool use."
- ✅ SHOULD TRIGGER — "Write the Python code for a CrewAI agent that handles client intake."
- ✅ SHOULD TRIGGER — "Implement an MCP server for our ticketing system."
- ❌ SHOULD NOT TRIGGER — "Design how I should structure my multi-agent system before I build it" → route to orchestrating-ai-agents.
- ❌ SHOULD NOT TRIGGER — "Connect n8n to Base44 using no-code" → route to workflow-automation.
When to Use / When NOT
Use when — "build an MCP server with OAuth that connects our CRM." "orchestrate a multi-LLM workflow with cost-based routing."
Do NOT use when — "explain the architecture patterns for multi-agent systems" — route to orchestrating-ai-agents.
AI Agents & Automation Workflow Skill v2.0
Build secure, modular, production-ready AI agents and automation workflows across the modern AI/automation ecosystem with MCP Protocol 2025 compliance, Claude Skills integration, and enterprise security patterns.
What's New in v2.0 (March 2026)
- MCP November 2025 Spec: Async operations, OAuth flows, stateless servers, capability declarations
- Security Hardening: April 2025 analysis compliance (prompt injection defense, tool sandboxing, consent flows)
- Framework Updates: LangGraph 1.0 stable, CrewAI multimodal + agentic RAG, market data
- Code Execution Patterns: On-demand tool loading for 100+ tool scenarios
- Claude Skills Integration: Description optimization, progressive disclosure, enterprise deployment
- Interoperability: Agent2Agent (A2A) protocol, cross-framework communication
Quick Start Decision Tree
Choose your approach based on requirements:
- Need direct Claude API integration? → MCP Server (see
references/mcp-servers.md) - Need no-code workflow? → Zapier/Make (see
references/tool-integrations.md) - Need multi-agent orchestration? → CrewAI/LangChain (see
references/agent-frameworks.md) - Need custom LLM routing logic? → Multi-LLM patterns (see
references/multi-llm-patterns.md) - Need web automation? → Playwright + MCP (see
references/tool-integrations.md) - Need 100+ tools with efficiency? → Code Execution + MCP (see
references/code-execution-patterns.md)
Core Workflow
Step 1: Requirements Gathering
Ask the user:
- Trigger: What initiates the workflow? (API call, webhook, schedule, manual)
- Data sources: Which tools provide input? (GitHub, Notion, etc.)
- LLM tasks: What requires AI reasoning vs. deterministic logic?
- Destinations: Where do results go? (Notion, GitHub, Slack, etc.)
- Security: Authentication requirements, secrets management, rate limits, consent flows
- Scale: How many tools? (determines if code execution pattern needed)
- Deployment: Where will this run? (local, Vercel, Cloudflare Workers, etc.)
Step 2: Architecture Selection
Use the appropriate pattern:
Pattern A: MCP Server (Claude-native tool integration, MCP 2025 spec compliant)
- Use when: Building reusable tools for Claude desktop/web/API
- Features: Async operations, OAuth flows, stateless design
- See:
references/mcp-servers.md - Init:
scripts/init_mcp_server.py
Pattern B: LangChain Agent (Python ecosystem, LangGraph 1.0 stable)
- Use when: Need tool calling with memory + Python integrations
- Features: Graph-based workflows, LangSmith observability
- See:
references/agent-frameworks.md§ LangChain - Init:
scripts/init_langchain_agent.py
Pattern C: CrewAI Workflow (Multi-agent collaboration, multimodal + agentic RAG)
- Use when: Need role-based agents working together
- Features: Independent from LangChain, native vector DB integration
- See:
references/agent-frameworks.md§ CrewAI - Init:
scripts/init_crewai_workflow.py
Pattern D: Multi-LLM Orchestration (Route tasks to best model)
- Use when: Need cost/speed/capability optimization
- See:
references/multi-llm-patterns.md
Pattern E: No-Code Integration (Zapier/Make)
- Use when: Non-technical stakeholders need to modify
- See:
references/tool-integrations.md§ No-Code Platforms
Pattern F: Code Execution + MCP (100+ tools, on-demand loading)
- Use when: Need efficient context management with large tool sets
- Features: Filesystem-based discovery, search_tools pattern
- See:
references/code-execution-patterns.md
Step 3: Security Baseline (April 2025 Compliance)
MANDATORY for all patterns:
- Input validation (never trust external data, prevent SQL injection)
- Secrets via environment variables (
.envfor local, secrets manager for prod) - Rate limiting (prevent abuse, per-user and per-endpoint)
- Auth/permissions (explicit scopes, least privilege)
- Error handling (no secret leakage in errors, structured cause→fix→retry)
- Prompt injection defense: Treat tool descriptions as untrusted
- Tool sandboxing: Implement least-privilege tool access
- Consent flows: Explicit user consent before tool invocation
Run: scripts/validate_workflow.py <path> to check security baseline
See: references/security-checklist.md
Step 4: Implementation
Build the workflow following the selected pattern. Use templates from assets/ as starting points:
assets/mcp-templates/- MCP server boilerplates (2025 spec compliant)assets/workflow-templates/- Zapier/Make workflow JSONsassets/agent-configs/- LangChain/CrewAI configurationsassets/deployment-configs/- Vercel/Cloudflare deployment configsassets/security-templates/- Consent flow implementations, input validationassets/code-execution-templates/- search_tools pattern, on-demand loading
Step 5: Testing
Test integration points before deployment:
- Use
scripts/test_integration.pyfor automated testing - Verify error handling (simulate API failures)
- Check auth (try with invalid credentials)
- Validate rate limits (burst test)
- Test edge cases (empty responses, malformed data)
- Security testing: Prompt injection attempts, unauthorized tool access
- Performance testing: Token consumption with large tool sets
Step 6: Deployment
Deploy according to pattern:
- MCP: Install locally or distribute .zip, register with claude.ai
- LangChain/CrewAI: Deploy to Cloud Run, Railway, Fly.io
- Vercel: Use
vercel deploywith edge functions - Cloudflare Workers: Use
wrangler publish - Enterprise: Organization-wide skill deployment (Dec 2025+)
See: references/deployment.md
Progressive Disclosure
Load reference files as needed:
references/mcp-servers.md- MCP server guide (2025 spec with async, OAuth, stateless)references/agent-frameworks.md- LangChain/CrewAI deep dive (2026 data with LangGraph 1.0, CrewAI multimodal)references/multi-llm-patterns.md- LLM routing strategiesreferences/tool-integrations.md- All tool APIsreferences/security-checklist.md- Security requirements (April 2025 compliance with prompt injection defense)references/code-execution-patterns.md- On-demand tool loading for 100+ tool scenariosreferences/claude-skills-integration.md- Skills architecture best practices (description optimization, enterprise deployment)references/interoperability.md- Cross-framework communication (A2A protocol, MCP coordination)references/deployment.md- Deployment guides (Vercel, Cloudflare, Railway, Docker, enterprise)
Quality Gates
Before marking any workflow complete, verify:
✅ All inputs validated (type checks, schema validation, SQL injection prevention)
✅ Secrets in environment variables only (no hardcoded tokens)
✅ Auth explicitly configured (scopes documented, least privilege)
✅ Rate limits implemented (per-user, per-endpoint, prevent API abuse)
✅ Error messages follow cause→fix→retry pattern
✅ Security baseline: Prompt injection defense, tool sandboxing, consent flows
✅ MCP 2025 compliance (if applicable): Async operations, OAuth flows, stateless design
✅ README includes setup + .env.example
✅ At least 1 happy-path test + 1 edge case test + 1 security test
✅ Deployment instructions provided
Version History
v2.0.0 (March 2026):
- ✅ Optimized description for 90% activation (added "Use when..." triggers with 20+ keywords)
- ✅ Added MCP November 2025 spec features (async, OAuth, stateless, code execution)
- ✅ Integrated April 2025 security analysis findings (prompt injection, tool sandboxing, consent)
- ✅ Updated framework comparison with 2025-2026 data (LangGraph 1.0, CrewAI multimodal/RAG)
- ✅ Added code execution patterns for 100+ tool scenarios
- ✅ Added Claude Skills architecture best practices (progressive disclosure, enterprise deployment)
- ✅ Added interoperability standards (A2A protocol, MCP coordination)
v1.0.0 (January 2026):
- Initial release with comprehensive automation workflow coverage
FAIL Conditions
Named failure modes for this skill. Each is a hard stop, not a warning.
- Passing unsanitized tool output back into the model as instructions → prompt injection via tool result — treat tool output as data, validate and escape before re-prompting.
- Placing secrets or OAuth tokens in agent context or traces → credential leak — inject secrets server-side at call time, never into prompt or state, and redact in observability.
- Executing tool calls with no sandbox or consent gate on consequential actions → unbounded blast radius — sandbox tool execution and require HITL approval for irreversible or spendy actions.
- An agent loop with no turn or cost ceiling → runaway token spend or an infinite loop — set a hard max-turns and budget guard and circuit-break on repeated failure.
- Stateful assumptions inside a stateless MCP server → corrupted multi-client state — hold no per-connection global state, key state by request or session.
- Shipping with no observability on tool calls and failures → silent degradation — log every tool call, its latency, and its errors, and wire to instrumenting-agent-telemetry.