Telemetry, security, and governance tooling for autonomous AI agents.

Your AI agent just approved a $50K purchase order. Can you prove why it did?

Agents loop, fabricate, and drift — and mostly, nobody's watching. We build tools to log what actually happened, catch what shouldn't, and give agents governable rules to act within. Not because we distrust agents — because trust requires evidence.

An independent R&D effort.

Static security analysis for AO/Lua processes. Catches auth bypasses, determinism violations, nil-guard gaps, unsafe JSON, and replay bugs before they hit chain. Tree-sitter parser, 25+ built-in checks, extensible with your own YAML rules. Runs as a CLI, MCP server, or GitHub Action.

Talk to AO/Arweave from Claude or Cursor: query processes, send messages, spawn processes, execute Lua, wallet-based signing. Five tools, zero config.

On-chain and off-chain tracking for autonomous agents: structured logging of every output, a confabulation detector, vocabulary-drift measurement (Jaccard), token accounting (actual vs requested), and stripped-context snapshots for comparing model state.

Small, honest experiments on how autonomous agents behave and govern.

• Governance as Computation — in a multi-agent commons dilemma, giving agents the pre-computed answer (not more information, not moral prompts) is what prevents collapse: 80% vs 0% survival. Preliminary (N=5), with full data and an honest limitations section.
• Agent Bootstrap — 160 words of flat navigational anchors beat 1,000 words of injected memory (174 trials).
• Movable Feast — LLMs recall fixed holidays but can't compute movable ones (Easter: 0% across models).

Independent research on trustworthy autonomy.