AEGIS - Autonomous Enterprise Graph Intelligence System

A self-hosted, auditable alternative to Glean + Devin + PagerDuty Autopilot, built 100% on LangChain.

Try it live: aegis-agent-api.vercel.app/ui - toggle between Demo and Live inference, watch the LangGraph supervisor route specialists in real time, and approve/reject HITL gates.

AEGIS takes a natural language operational request - "Why is checkout latency spiking in us-east?" - and autonomously plans, delegates to specialist sub-agents, retrieves from hybrid knowledge bases, executes tools, hits human-in-the-loop gates, and posts a fully traced, evaluated, and auditable result.

---

Live Demo

Open aegis-agent-api.vercel.app/ui and click Run AEGIS.

What you'll see:

• Real-time LangGraph visualizer - labeled specialist pills around the supervisor, animating Supervisor → SRE Analyst → Knowledge → Coder → [HITL] → Evaluator → Communicator. Names stay inside the node so they do not clip on a phone.
• Specialist roster - idle / live / done for each agent in the current run
• Streaming specialist cards - each agent's findings appear as they execute, with confidence scores and artifact counts
• Human-in-the-Loop gate - the Coder produces a patch, pauses for your approval, then the Evaluator and Communicator complete the flow
• Demo / Live toggle - Demo mode runs an instant simulation; Live mode connects to the real LangGraph with your API keys
• Live info panel - step count, confidence %, artifact count, and elapsed time update in real time
• LangSmith traces - one-click link to the full trace for every run

Screenshots

Public demo threat model: SECURITY.md. Demo/sim is the UI default; Live inference unlocks when an LLM key is present (LIVE_MODE=false forces demo). Optional PUBLIC_RUN_TOKEN. Rate-limited. Not bank-grade.

---

Architecture

[FastAPI dashboard / LangGraph Studio] <-SSE-> [LangServe FastAPI /api]
                                        |
                              [LangGraph Supervisor]
                   /     |      |       |       |      \
            Researcher Coder  SRE   Knowledge Comm  Evaluator
               |         |     |        |
         Tavily/Arxiv  E2B  Prometheus  PGVector Hybrid RAG
                                        |
                                [Postgres + PGVector + Redis]
                                        |
                              [LangSmith Traces / Evals / Prompt Hub]

Feature Matrix - Full LangChain Ecosystem

Product	Used For
langchain-core	LCEL everywhere, structured output Pydantic v2, fallback LLM router
langgraph	Supervisor + 6 subgraphs, PostgresSaver, interrupt() HITL, astream_events
langsmith	Tracing, Prompt Hub (aegis/supervisor_router), Evals, Feedback API
langserve	FastAPI /invoke, /stream, /threads/{id}/resume, OpenAPI playground
RAG	MultiQuery → Cohere Rerank → LLM Grader → HyDE, PGVector + BM25 hybrid
Tools (14)	Tavily, Code Executor, Postgres, GitHub, Slack, Browser, Prometheus, Runbook, Arxiv, Wikipedia, Email, Calendar, FS, Memory

7 Agentic Loops - All Implemented

1. Perception-Plan-Act-Reflect
2. Supervisor-Worker Hierarchical
3. RAG Self-Correction
4. Tool-Use ReAct + Self-Heal
5. Human-in-the-Loop Interrupt
6. Evaluation-Driven Self-Improvement
7. Memory Consolidation

All visible in LangSmith with custom metadata.

---

Quickstart

Vercel (recommended - zero config)

1. Fork this repo
2. Import into Vercel
3. Set root directory to apps/api
4. Add GOOGLE_API_KEY (or OPENAI_API_KEY) as an environment variable
5. Deploy - visit /ui for the live dashboard, /docs for the API playground

Without API keys the UI gracefully falls back to Demo mode (instant simulation).

Docker (local / self-hosted)

cp .env.example .env
docker-compose -f infra/docker-compose.yml up --build

• Dashboard: http://localhost:8000/ui
• API playground: http://localhost:8000/docs
• LangGraph Studio: langgraph dev

API Endpoints

Endpoint	Method	Purpose
/ui	GET	Live dashboard (SSE, visualizer, HITL)
/stream	POST	Streaming inference (SSE)
/invoke	POST	Single-shot inference (JSON)
/threads/{id}/resume	POST	Resume after HITL (JSON)
/threads/{id}/resume/stream	POST	Resume after HITL (SSE)
/health	GET	Graph status, key availability
/docs	GET	OpenAPI / Swagger playground

---

Why This Proves Senior+ AI Engineering

• Agentic Loops: 7 explicit loops, not chains
• LangGraph HITL: interrupt() / Command(resume=...), PostgresSaver
• LangSmith Evals/Prompt Hub: 3 datasets, LLM-as-judge, CI gating faithfulness > 0.82
• Hybrid RAG: MultiQuery + Compression + Grader + HyDE
• Multi-agent Supervisor: 6 specialists, tool-use ReAct
• Production Observability: OpenTelemetry → LangSmith, run metadata
• Vercel Serverless: graceful degradation, SSE streaming, version-agnostic chunk handling

Repo Structure

aegis/
├── apps/api/              # LangServe FastAPI + live UI
├── packages/aegis_graph/  # Supervisor + 6 subgraphs
├── packages/tools/        # 14 production tools
├── packages/rag/          # Ingestion / retriever / vectorstore
├── packages/memory/
├── packages/evals/
├── infra/docker-compose.yml
├── tests/
└── scripts/run_evals.py

Evals

python scripts/run_evals.py

Writes evals/reports/latest.md. Public CI has no LANGCHAIN_API_KEY, so that job writes a mock report and exits 0 - it does not measure live LangSmith faithfulness. With the key set, datasets aegis_rag_qa, aegis_tool_use, and aegis_incident_triage run against project aegis-production; the intended production threshold is faithfulness ≥ 0.82.

CI itself fails on ruff (real errors), mypy on tools/evals/tests, and pytest (graph compile, RAG loop, tool guards, /health /ui /stream smokes). pip-audit is informational and does not fail the job on LangChain majors.

Environment Variables

Var	Purpose
GOOGLE_API_KEY	Gemini LLM (primary)
OPENAI_API_KEY	OpenAI fallback
ANTHROPIC_API_KEY	Coding fallback
LANGCHAIN_API_KEY	LangSmith tracing
LANGCHAIN_TRACING_V2=true	Enable tracing
DATABASE_URL	Postgres + PGVector
REDIS_URL	Short-term memory
TAVILY_API_KEY	Web search

All optional - fake models/fallbacks keep Vercel deploy green even without keys.

---

MIT License - Built with LangChain, LangGraph, LangSmith