Initial commit: PRD v0.2, PROGRESS, knowledge base, PoC

poc/README.md

@@ -0,0 +1,66 @@
+# FlowPilot Agent Orchestration PoC
+
+This proof-of-concept demonstrates the core agent orchestration ideas from the FlowPilot PRD:
+
+## Core Concepts Demonstrated
+
+1. **HR Manager Agent**: Decomposes high-level tasks into atomic tasks and creates executor agents
+2. **Experience Manager Agent**: Logs executions, provides context, and maintains a knowledge base
+3. **Atomic Task Decomposition**: Breaking down complex tasks into single-agent-completable units
+4. **Agent Lifecycle Management**: Dynamic creation of executor agents for specific task types
+
+## How to Run
+
+```bash
+python3 agent_poc.py
+```
+
+## Output
+
+The PoC processes three sample high-level tasks from the project management handbook:
+- Creating a project charter
+- Identifying software development risks
+- Creating a project schedule
+
+For each task, it shows:
+- How the HR Manager decomposes the task into atomic tasks
+- How executor agents are created for each atomic task type
+- How the Experience Manager logs executions and provides context
+- The final results and execution logs
+
+## Next Steps for Development
+
+1. **Replace simulation with actual LLM calls** - Integrate with OpenAI/Anthropic/etc APIs
+2. **Add real task execution** - Instead of mock results, have agents actually perform tasks
+3. **Implement persistence** - Store agent configurations, execution logs, and knowledge base in database
+4. **Add feedback loops** - HR Manager improves based on execution scores
+5. **Integrate with Feishu** - Build the actual frontend and Feishu bot interfaces
+
+## Files
+
+- `agent_poc.py`: Main PoC implementation
+- `poc_results.json`: Detailed execution results from the last run
+- `.gitignore`: Git ignore file
+
+## Design Notes
+
+This PoC focuses on demonstrating the orchestration logic rather than actual AI execution.
+In a real implementation:
+- The `_simulate_execution` method would call actual LLM APIs
+- Agents would have specific prompt templates for their task types
+- The Experience Manager would use vector embeddings for context retrieval
+- Task decomposition would be more sophisticated (possibly using LLMs themselves)
+
+## Relation to PRD
+
+This PoC validates the core architectural concepts from PRD v0.2:
+- Dual agent management (HR Manager + Experience Manager)
+- Atomic task decomposition
+- Agent lifecycle management
+- Context sharing and knowledge base
+
+It does not yet cover:
+- Feishu integration
+- Full PMBOK process coverage
+- Production-level error handling and scaling
+- Actual LLM provider integration