Copilot instructions

1---
2applyTo: "**/crewai/**/*.py,**/crew/**/*.py,**/agents/**/*.py"
3---
4 
5# CrewAI Development Standards
6 
7## Agent Design
8 
9- Define `role`, `goal`, and `backstory` for every agent — all three are required for effective specialisation
10- Keep each agent's role narrowly focused — one clear responsibility per agent
11- Assign only the tools each agent needs — do not give all tools to all agents
12- Use `verbose=True` during development; disable or redirect output in production
13 
14## Task Design
15 
16- Every task must have a `description`, `expected_output`, and assigned `agent`
17- Use `context=[prior_task]` to chain outputs between tasks
18- Define `expected_output` with enough specificity to guide the agent's output format
19- Prefer Pydantic model outputs for structured data tasks
20 
21## Process Selection
22 
23- Use `Process.sequential` for ordered, dependent task chains
24- Use `Process.hierarchical` only when a manager agent must dynamically route to workers
25- Always define `max_rpm` or `max_execution_time` to prevent runaway crew execution
26 
27## Tool Integration
28 
29- Subclass `BaseTool` for all custom tools; annotate `_run` with type hints
30- Handle tool errors gracefully — return error messages as strings, never raise unhandled exceptions
31- Enable tool caching (`cache_function`) for expensive or external-calling tools
32 
33## Safety
34 
35- Always test crew output with representative inputs before production deployment
36- Monitor token usage across all agents in a crew — multi-agent costs accumulate quickly
37- Implement output validation before passing crew results to downstream systems
38