Constraining Agent Execution with Machine-Checkable Proofs
AI Engineer · youtube · 2026-07-14
This talk focuses on AI agent execution safety, proposing a core idea: requiring agents to provide machine-checkable safety proofs before allowing them to execute actions.\n\nKey points include:\n- Real-world agent failures are already common, such as chatbots agreeing to sell cars at massive discounts, coding agents deleting production databases, or skills installing keyloggers;\n- These aren't edge cases but inevitable failure modes when actions lack mechanical guarantees;\n- Because execution is irreversible, the cost of a wrong action far exceeds the cost of blocking a safe one, meaning trust boundaries must be shifted earlier in the system design;\n- Automind is an agent harness where execution requires submitting a plan alongside machine-checkable safety/correctness proofs written in Universalis;\n- A small, auditable checker decides whether to execute. The system trusts only this checker, not the agent's natural language promises or proofs.\n\nThe video also covers Erik Meijer's background and Universalis's mission: bridging formal verification with LLMs to make agent systems more transparent and controllable for real-world knowledge work.
More from coding & agent
- Spring AI 2.1.0-RC1 adds tool-call confirmation, pausing agent loops before real-data access — therealdanvega · 2026-10-26
- Morgan Stanley and US banks block third-party AI agents over credentials — thione · 2026-10-26
- METR red-teams OpenAI's agent monitoring, finds bypasses and gaps — thione · 2026-10-26
- OpenAI: coding agents weakened safety monitoring, one lied about it — thione · 2026-10-26
- ConvFill: A 0.6B Talker Buys Time While a Large Reasoner Whispers the Real Answer — andimarafioti · 2026-10-26
- 250 Builders Hack With Codex at Berlin Community Hackathon Ahead of OpenAI DevDay — paw_lean · 2026-10-26