finds.dev

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。

View on GitHub

★ 302Pythonupdated Sep 2026

// the take

AegisOps is a Chinese-language AIOps workbench for internal enterprise systems: scheduled health checks feed a LangGraph/Deep Agents diagnosis pipeline that gathers evidence (metrics, logs, traces, change history) and drafts a ticket for human approval before any remediation runs. It's aimed at platform/SRE teams who want an agent that investigates and proposes, not one that has shell access and acts on its own.

The approval boundary is actually enforced in the architecture, not just policy: the agent only reads evidence through read-only MCP tools within a fixed asset/time window, and destructive actions require a separate human-approved step with recorded parameters and an expiry. Resumability is tested, not assumed — checkpoints, leases, and idempotent writes were validated across 20 interrupted-process recovery runs with no duplicate tickets. The benchmark methodology is unusually transparent for this space: 24 paired runs of the same DeepSeek model over 12 public incident replays (RCAEval/Loghub), with token counts, call counts, and median latency reported, plus raw run logs checked into eval/results.

The authors' own docs admit Docker Compose has only passed config validation — container build and run are explicitly untested, so anyone not on Apple Silicon is the first to find out if it actually works. The pgvector-based knowledge retrieval, one of the headline capabilities, is flagged as unverified for embedding provider quality; default search is still plain keyword matching. Only DeepSeek has been validated for the live tool-calling path, so swapping in another model is unproven. The reported speed/token wins (33.7% faster, 53.7% fewer tokens) come from a 12-case public benchmark the authors themselves caveat as not representative of production, and the comparison baseline tuning (forensic rounds cut from 8 to 4) makes it as much a prompt-engineering result as an architecture one.

We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
302
Forks
48
Language
Python
Last commit
10 Sep 2026

// nearby

If AegisOps isn't quite it.

All repos →