HOLD

What happens when an AI agent says “done” but has no proof? AI agents increasingly act in the real world: booking travel, making purchases, processing payments and operating business systems. But intelligence should not grant itself permission to act—or certify its own success. HOLD separates intelligence, authority and evidence. Imagine an AI agent booking a hotel. The reservation request reaches the provider, but the confirmation disappears. Did it succeed? Retrying risks a duplicate booking. Declaring success without evidence misleads the user. Stopping permanently defeats autonomy. HOLD introduces a third path: governed recovery. A human defines the mission and its Habilitation: permitted actions, limits, conditions and escalation rules. Meta Llama reasons and proposes actions. HOLD's deterministic kernel independently admits, refuses or holds execution. When an outcome becomes uncertain, HOLD records UNKNOWN, investigates the existing operation and requests bounded human intervention when necessary. Only correlated evidence can establish completion. Built during the hackathon We implemented a working Python prototype featuring: Habilitation Profiles and deterministic enforcement. Uncertainty reconciliation and bounded human escalation. Hash-chained evidence journal and offline replay. Interactive mission demo and Meta Llama 3.2 3B integration. Reproducible experimental framework and evidence exports. Measured results We evaluated 99 synthetic trajectories across three conditions using Meta Llama 3.2 3B. 66 recorded violation categories in the ungoverned baseline. 0 detected executed violations across 33 HOLD-governed runs. 8 legitimate missions completed with correlated evidence. 19 duplicate-booking attempts refused. An external AI-assisted audit examined all 99 trajectories, with qualified support for the results. These findings demonstrate feasibility in a controlled environment, not universal safety. The guided demo uses a synthetic hotel and simulated human relay. Meta Llama's actual behavior is documented in recorded experiments. WhatsApp Cloud API relay support is implemented but not live-validated. Product vision and Meta opportunity HOLD is not a booking application. It is an emerging execution-governance layer for autonomous AI. Our vision is interoperable Habilitation Profiles: reusable governance specifications across models, tools and runtimes, while execution authority remains independently verified. The commercial opportunity includes developer SDKs, execution-governance APIs, enterprise assurance and industry-specific profiles. For Meta, HOLD could help developers deploy Llama-powered agents in more consequential workflows, accelerating enterprise adoption without relying on perfect model behavior. HOLD was implemented during the hackathon, building on earlier TENIR/REG research. Cross-runtime interoperability remains future work. Intelligence proposes. Authority governs. Evidence establishes.