Repository navigation
ACE v1.0 — First Release
Agent Certification Engine is now generally available.
ACE is an end-to-end platform for evaluating autonomous AI agents under real infrastructure chaos. It injects faults into a Kubernetes cluster, lets an AI agent attempt autonomous remediation, collects the full LLM trace, and produces a statistically rigorous 12-section certification report.
What's included
Certification pipeline
- 4-phase Python pipeline: fault bucketing → metrics extraction → statistical aggregation → certification report assembly
- LLM Council: k independent judges + meta-reconciliation to eliminate single-model bias
- Optional Phase E: H01–H09 statistical hypothesis tests (BCa bootstrap, Kruskal-Wallis, CUSUM drift detection) for n ≥ 30 runs
AgentCert control plane
- LitmusChaos ChaosCenter v3.0.0 with ACE-specific packages: agent registry, app registry, fault studio, observability, certification orchestrator
- Full GraphQL API (queries, mutations, subscriptions) with JWT RBAC
- React 17 + TypeScript frontend with Chaos Studio, Agent Hub, App Hub
Agents evaluated
flash-agent— FLASH-methodology SRE ITOps agent (custom ReAct loop + MCP)
Infrastructure
- One-command setup (
./scripts/setup.sh --deploy) across three cluster routes: existing cluster, fresh KinD, cloud (AKS/EKS/GKE) - Helm chart deploying the full platform stack (~36 Gi storage, 16 services)
- LiteLLM unified gateway routing to Azure OpenAI, Gemini, OpenRouter, or local Ollama
- Langfuse OTEL trace storage for all LLM calls
- Shared-host isolation — every container, volume, network, and cluster name is instance-scoped to prevent cross-checkout collisions
Quick start
git clone --recurse-submodules https://github.com/AgentCert/ace-monorepo
cd ace-monorepo
./scripts/setup.sh --deploy