GreatAgent — LangChain + Bedrock Demo

This repo now demonstrates a multi-agent workflow powered by Claude 3.5 Sonnet on AWS Bedrock, LangChain tooling, a Valyu search tool, and LangSmith tracing. The backend builds a mini orchestrator that walks through Intake → Task Decomposition → Workflow planning → Research → Analysis → Validation → Output, and the React dashboard streams each step while also showing LangSmith traces.

Architecture

• agent_core/bedrock_client.py — thin wrapper around Bedrock's Claude 3.5 Sonnet (falls back to demo text if AWS creds are missing).
• agent_core/pipeline.py — LangChain-style helper that prompts Claude for each agent stage and uses Valyu for search.
• controller/server.py — FastAPI controller that runs the pipeline, emits SSE events, fetches LangSmith traces, and exposes configuration endpoints.
• controller/langsmith_client.py — pulls recent traces for the dashboard.
• greatagent-ui/ — React + Vite + ReactFlow dashboard with a LangSmith trace panel, pipeline editor, and live stream view.

Configuration

Create or edit .env in the repo root (a template is already checked in). Important variables:

Variable	Purpose
AWS_REGION, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY	Needed to call Claude on Bedrock. When omitted, the app runs in demo mode with placeholder responses.
CLAUDE_MODEL	Bedrock model ID (default anthropic.claude-3-5-sonnet-20240620-v1:0).
VALYU_API_KEY, VALYU_API_URL	Credentials for the Valyu search API. Demo snippets are used when missing.
LANGCHAIN_TRACING_V2, LANGCHAIN_ENDPOINT, LANGCHAIN_API_KEY, LANGCHAIN_PROJECT	Standard LangSmith env vars. When set, every LangChain call automatically pushes traces.
LANGSMITH_API_KEY, LANGSMITH_PROJECT, LANGSMITH_API_URL, LANGSMITH_DASHBOARD_URL	Used by our /traces endpoint and to build shareable trace links.

You can still edit per-agent prompts/model IDs from the dashboard (Agent Settings panel) — they’re stored in agent_settings.json via the existing settings store.

Running the demo

1. Install deps (Python 3.11+ recommended):

   pip install -r requirements.txt
   cd greatagent-ui && npm install && cd ..

2. Start the backend:

   uvicorn controller.server:app --reload --port 8077

3. Start the frontend:

   cd greatagent-ui
   npm run dev

4. Open the dashboard (default http://localhost:5173). Enter a query and click Start. Each stage streams into the UI, and the LangSmith Traces panel links to the latest trace (demo URLs when LangSmith isn’t configured).

LangSmith dashboard data

• Every run emits a trace event with the shareable URL from AgentPipeline.start_trace().
• The new /traces endpoint aggregates recent runs via the LangSmith API. The React dashboard polls it whenever a run completes and lists the latest trace cards.

Notes

• Claude / Valyu calls fall back to deterministic demo output when credentials are missing, so you can preview the UX without real keys.
• The ReactFlow canvas is still editable and prettier than before; save layout + plan to customize future runs.
• Agent API settings are optional in this demo but remain editable so you can point specific stages at bespoke microservices later.