Foreman

Foreman is the Kubernetes-native control plane for agentic workloads that ships as an opt-in add-on to LLMKube. It dispatches coder, verifier, and reviewer agents across a heterogeneous fleet of local LLM nodes. We assess it: the coder/verifier/reviewer split is a sound shape for agentic coding, but it inherits Kubernetes complexity.

Blurb

Foreman is the Kubernetes-native control plane for agentic workloads that ships as an opt-in add-on to LLMKube. It dispatches coder, verifier, and reviewer agents across a heterogeneous fleet of locally-hosted LLM nodes.

Summary

Foreman models agentic coding as Kubernetes CRDs (Workload, AgenticTask, Agent, FleetNode) in its own API group and Helm chart, so it drops onto an existing LLMKube/K8s surface rather than fighting it. A Workload (intent + repo + issues) is decomposed by a reconciler into coder, verifier, and reviewer tasks; capability-aware scheduling claims each task on a fleet node whose hardware fits.

The v0.1 pipeline is deliberately linear: coder edits and pushes a fork branch, verifier runs a gate command (e.g. make lint test) and emits GATE-PASS/GATE-FAIL, and reviewers score the diff and emit APPROVE/REQUEST-CHANGES/REJECT. Reviewer ensembles and a coder escalation tier (fast model first, heavier model for capability failures) are first-class.

Why the coder/verifier/reviewer split matters: if you are going to use an agent for coding, you should think in those terms. A separate non-agentic verifier gate grounds the coder’s claims, and a reviewer catches what the gate misses. In-process frameworks (CrewAI, LangGraph, AutoGen) don’t solve fleet dispatch; Foreman sits between the inference engine and the application-layer agent framework.

When to use: you already run Kubernetes for inference and want reproducible, self-hosted agentic pipelines with humans in the loop. When to skip: single-machine or light setups where a simpler harness like Claude Code is enough. v0.1 lacks DAGs, best-of-N, and an autonomous planner (v0.2+).

Details

Four CRDs: Workload (user intent), AgenticTask (dispatchable unit), Agent (reusable role: system prompt, tool whitelist, model endpoint, required capability), and FleetNode (cluster-scoped, self-registered host capability). Inference is local by default against your own InferenceServices (vLLM, llama.cpp, mlx-server); no cloud egress unless you opt in.

The verifier gate runs as a Kubernetes Job and is non-agentic. Reviewers emit structured findings against an A-through-H checklist. Escalation fires only on capability failures (NO-GO or coder-gate failure), never on INCOMPLETE or stuck-loop, since a bigger model won’t fix a give-up. It assumes OpenAI tool_calls compatibility across inference endpoints.