Skip to content

Which version do I want?

You're reading the v2 docs — the lean baseline. This repository actually holds four generations of KubeIntellect side by side. They are not a version history where the newest replaces the last one; each is a self-contained re-architecture, kept on purpose because the differences between them are the research result.

Just want the diagnosis-and-approval workflow, today? v2 (this docs site) is a fully working, minimal way to do that.

Want the current, actively-developed platform — memory, autonomy, predictive detection? That's v4.

v4 is where feature work happens now. It isn't published under this docs site yet; see the GitHub repo for its own docs.

The four generations, at a glance

v1 v2 (this site) v3 v4
What it is Capability-maximal origin Lean baseline Framework-delegated Platform
Agents Supervisor + 13 specialised ReAct agents 1 coordinator, on-demand 4-subagent RCA Coordinator + sub-agents via deepagents Lean coordinator + feature-flagged layers
Tools ~100+ runtime-synthesised tools 4 guarded tools (run_kubectl, run_helm, query_prometheus, query_loki) Same 4-tool core, virtual-filesystem planning Same core + sensorium, memory, flight recorder
Safety gate Per-agent checks 7-layer kubectl safety guard Inherits v2's guard Inherits v2's guard + RBAC tiers
Status ❄️ Frozen — the published architecture 🔸 Baseline — bug fixes and docs only 🔸 Experimental — bug fixes and docs only ✅ Current — all feature work
Cited by JGC 2026 paper Fair-comparison campaign — —

Lineage: v1 (capability-maximal) → simplify → v2 (lean, measurable) → reframe → v3 (framework-delegated) → productionize → v4 (platform).

v1
Frozen
Capability-maximal
13 specialised ReAct agents, ~100+ runtime-synthesised tools. The architecture the paper describes.
v2
You are here
Lean baseline
One coordinator, 4 guarded tools, a 7-layer kubectl safety guard. Simpler — and the finding was that simpler also scored better.
v3
Experimental
Framework-delegated
v2's behaviour re-expressed through deepagents — a virtual filesystem plus task delegation.
v4
Current
Platform
The lean coordinator plus feature-flagged layers: sensorium, memory hierarchy, flight recorder, autonomy ladder.

Why keep all four instead of just shipping the latest?

Because the finding is the comparison. The interesting result of this project was that the 13-agent, ~100-tool system (v1) was worse than a lean coordinator (v2) — on quality and on cost. Deleting v1 would delete the evidence for that claim. Each version has its own Makefile, docs/, tests/, and packaging; they are independent snapshots of a design lineage, not duplicated debt.

See v4 — the current platform — in action

The two videos below were recorded against v4, not against this v2 baseline. v4 adds a memory hierarchy, an autonomy ladder, and a sensorium/detector engine that v2 does not have — so treat these as "where the project is headed," not as a demo of what you get by installing v2.

KubeIntellect v4 demo thumbnail KubeIntellect v4 — full demo, 8 minv4
KubeIntellect v4 architecture animation thumbnail v4 architecture, animated — 38sv4

Do not mix them

Each version has its own Makefile, docs/, tests/, and .env.example. What they share is the infrastructure underneath — one Kind cluster, one observability stack (Prometheus + Grafana + Loki) — managed from the repository root.

Where to go next