HN Showcase
Back to browse
GitHub Repository

Constrained autonomy runtime for AI agents.

1 starsRust

HELmR – A runtime control layer for autonomous agents

by systems_arch·Mar 9, 2026·2 points·1 comment
Visit ProjectView on HN

AI Analysis

●●SolidBig BrainBold Bet

Deterministic agent governance with capability tokens beats probabilistic guardrails.

Strengths
  • •Agents cannot execute actions directly — all must pass through authorization airlock.
  • •Mission budgets and tomb-state blocking prevent runaway agent behavior.
Weaknesses
  • •Only 8 commits and zero stars — very early stage with unclear integration path.
  • •No documentation on how this connects to LangChain, AutoGen, or other frameworks.
Category
AI/ML
Target Audience

Developers building autonomous agents with safety requirements

Similar To

LangSmith · Arize Phoenix · Cognition Guardrails

Post Description

I built HELmR as an experiment in runtime governance for autonomous agents.

Most agent frameworks allow agents to execute actions directly against systems (filesystem, APIs, shell commands, etc). That means governance is optional and enforcement is outside the execution path.

HELmR takes a different approach: agents cannot execute actions directly. Every action must pass through HELmR authorization and a controlled execution airlock.

The system enforces:

• mission budgets • capability tokens • deterministic authorization • controlled execution • agent termination with a tomb registry

The goal is to explore whether agent governance should look more like infrastructure (similar to IAM or API gateways) rather than relying on probabilistic guardrails.

Curious what people think about this architecture.

Similar Projects

AI/ML●●Solid

Axon – Open-source agentic AI with approval gates (Apache 2.0)

AXON's core move — surfacing every tool call with a low/medium/high risk label and requiring Allow/Reject/Allow-for-session — is a practical, under-explored control for agentic AI and immediately useful for regulated environments. The repo pairs that UX with a React UI, FastAPI backend, Docker sandboxed code execution and multi‑LLM integrations (Ollama/Claude/OpenAI), so it feels like a real starter stack rather than a sketch. Biggest unknowns are adoption and ecosystem of plugins/skills — the idea is solid, but it needs community momentum to matter.

Niche GemSolve My Problem
NeuroVexon
124mo ago