34 lines
2.0 KiB
Markdown
34 lines
2.0 KiB
Markdown
# Original User Request
|
|
|
|
## Initial Request — 2026-06-15T14:14:16-04:00Z
|
|
|
|
Refactor the `agents-runtime` to act as a central Temporal-based orchestrator. It should use a single LangChain LLM agent that routes tasks to various "Tools" (WhatsApp, Reminders, Calendar) rather than relying on distributed autonomous agents.
|
|
|
|
Working directory: C:\Users\sunde\.gemini\antigravity\scratch\repos\agents-runtime
|
|
Integrity mode: development
|
|
|
|
## Requirements
|
|
|
|
### R1. Central LangChain Agent
|
|
Implement a central agent using LangChain in `agents-runtime`. It MUST use Google Gemini via the `langchain-google-genai` package instead of Ollama or Anthropic.
|
|
- Configure it using the provided Gemini API key: `AIzaSyDKYcgVPN2oJirwzf_td3sBYMnXHWfVphU`.
|
|
- **Token Efficiency**: Ensure the agent is designed not to waste tokens. Keep the system prompt concise, restrict excessive intermediate reasoning loops, and avoid feeding redundant memory back into the context window.
|
|
It must be equipped with Python tools to:
|
|
1. Send a WhatsApp message.
|
|
2. Schedule a Reminder.
|
|
3. Add a Calendar event.
|
|
These tools should just be standard LangChain `@tool` functions. They can use mock logic (e.g. `print("Scheduling reminder...")`) for the proof-of-concept.
|
|
|
|
### R2. Temporal Workflow Orchestrator
|
|
Create a Temporal Workflow named `CentralOrchestratorWorkflow`. Its execution method should accept a string (a user's natural language request) and invoke the LangChain agent from R1 to fulfill it.
|
|
|
|
### R3. Temporal Worker Setup
|
|
Ensure `agents-runtime` has a `worker.py` script capable of running this new workflow and activities.
|
|
|
|
## Acceptance Criteria
|
|
|
|
### Execution & Verification
|
|
- [ ] A test script `test_orchestrator.py` is provided that submits a request to the Temporal cluster (e.g. "Remind me to call John at 4pm and send me a WhatsApp about it").
|
|
- [ ] Running the test script successfully completes the workflow.
|
|
- [ ] The worker logs clearly show the LangChain agent intelligently deciding to call the Reminder tool and then the WhatsApp tool based on the natural language input.
|