emma-sue/AegisOps

AegisOps企业信息系统智能运维平台:自动巡检、Agent 故障排查、证据工单与审批处置。LangGraph · Deep Agents · MCP;附可复现实测记录。

What it is

AegisOps (shown in the repo as 企业信息系统智能运维 Agent) is an open‑source AI‑powered operations console for Chinese‑language enterprise IT environments. It ties together automated health‑checks, alert de‑duplication, an LLM‑driven “Agent” that gathers evidence from logs, metrics and change records, and a ticket‑workflow that lets human operators review the evidence, approve actions and close the incident.

The core idea is to let a language‑model‑based agent explain why a service is misbehaving, automatically collect the relevant data, and produce a draft work‑order that humans can verify and execute. The system can run in a fully‑demo mode that uses fixture data (no API keys) or in a live mode that talks to a real LLM (e.g., DeepSeek‑flash) for production use.


Key capabilities

Capability How it works
Scheduled inspections & alerts 8 built‑in health‑check rules run on a timer, generate alerts, de‑duplicate them and keep the raw alert data for later review.
LLM‑driven fault investigation A LangGraph workflow orchestrates Deep Agents (LLM agents). The agents load “Skills” on demand, read read‑only data via the MCP SDK, and produce a structured diagnosis report.
Evidence collection & ticket drafting For a chosen asset and time window the system stores raw evidence (logs, metrics, call‑chains) together with a hash. The diagnosis report links to this evidence and a work‑order draft can be assigned, annotated and later approved.
Knowledge reuse Documentation is indexed per organization/system/version. Users can search by keyword or (optionally) via pgvector semantic search. Approved cases and skills are curated for future reuse.
Human approval & safe remediation Work‑order approval binds the operation parameters, version and validity period. Check‑points and idempotent writes let the remediation be paused/resumed safely.
Progress & resource monitoring Server‑Sent Events (SSE) stream task progress. The platform caps the number of evidence‑gathering rounds, LLM calls and token usage, logging cost and latency.

Architecture & tech stack

  • Agent layer – LangGraph for workflow orchestration, Deep Agents for LLM reasoning, custom Skills and the MCP SDK for controlled tool access.
  • Data & task layer – PostgreSQL (with pgvector for optional semantic search) stores assets, tickets and knowledge; Redis Streams handle asynchronous task queues and checkpointing.
  • Services & UI – Spring Boot (Java) serves business logic, permissions and ticket APIs; FastAPI (Python) runs the agent workers; Vue 3 provides the Chinese‑language web console.
  • Dev / test tooling – Docker Compose for multi‑service orchestration, pytest & JUnit for unit tests, Playwright for end‑to‑end UI tests.

Typical workflow (visualised in the README)

  1. Detect – A scheduled inspection or a manual trigger raises an alert (e.g., “file‑processing slowdown”).
  2. Scope – The operator selects the affected asset and a time window; the system gathers raw evidence.
  3. Investigate – The LLM agent runs, consulting the evidence, dependency graph and any relevant procedural documents, then produces a diagnosis with candidate root causes and “to‑verify” items.
  4. Draft ticket – The diagnosis is attached to a work‑order draft that can be assigned, commented on and enriched with additional evidence.
  5. Approve – An authorized reviewer approves the remediation plan, which may include an isolated fault‑injection rehearsal.
  6. Execute & verify – The remediation runs, the system checks business‑level outcomes, and the final report is archived for future reference.

How well does it work?

The authors benchmarked the system on 12 public incident cases (24 paired runs) using the DeepSeek‑flash model:

  • Median diagnosis time dropped from 23.5 s to 15.6 s (‑33 %).
  • Total tokens per batch fell from 1.08 M to 0.50 M (‑54 %).
  • Model calls reduced from 108 to 60 (‑44 %). They also performed 20 interruption‑recovery tests where the agent correctly resumed work without duplicating tickets.

    Note: Benchmarks use replayed public data and a local demo environment; real‑world performance may vary.


Getting started (quick‑start)

# 1. Clone the repo
git clone https://github.com/emma-sue/-Agent-.git
cd -Agent-
cp .env.example .env   # edit if you want live LLM mode

# 2. Run a demo (macOS Apple Silicon only)
make demo-local   # downloads a self‑contained runtime and starts all services

# 3. Or use Docker Compose for other platforms
docker compose --env-file .env --profile live-lab up --build -d --wait

# 4. Open the UI
open http://127.0.0.1:5173   # login with user "ops-demo" and the demo password from .env

The demo ships with AI_MODE=fixture, so no API keys are required. To switch to a real model, set AI_MODE=live and fill in CHAT_BASE_URL, CHAT_MODEL and CHAT_API_KEY (e.g., DeepSeek credentials).


Who might use it?

  • Enterprise IT ops teams that need a unified console for monitoring, automated root‑cause analysis and ticketing.
  • SREs looking to prototype LLM‑assisted RCA (Root‑Cause Analysis) workflows without building the stack from scratch.
  • Developers of internal tooling who want to reuse the LangGraph‑based agent pattern for other domains (security, compliance, etc.).

Documentation & community

  • Docs hub – docs/INDEX.md links to running guide, API spec, code‑reading guide and benchmark details.
  • Issue tracker – https://github.com/emma-sue/-Agent-/issues for bugs, feature requests and deployment help.
  • Upstream components – The project reuses Deep Agents, LangGraph and the MCP SDK; see docs/upstream-reuse.md for licensing and contribution notes.

TL;DR

AegisOps is a Chinese‑language, open‑source AI operations platform that combines scheduled health checks, a LangGraph‑orchestrated LLM agent, evidence‑preserving diagnostics and a ticket‑approval workflow. It can run fully offline for demos or connect to a real LLM (e.g., DeepSeek) for production use, and it ships with Docker‑Compose or a native macOS demo runtime for quick evaluation.

Related

  • Project
  • Project
  • Project
  • Project