Toward a Science of Intent: Closure Gaps and Delegation Envelopes for Open-World AI Agents
Summary
This conceptual framework introduces "intent compilation" for open-world AI agents, addressing why capable models often fail deployment in institutional settings despite their problem-solving abilities. The authors propose transforming partially specified human purpose into inspectable artifacts that bind execution, distinguishing between closed-world solvers and open-world agents. They formalize residual openness as a "closure-gap vector" across semantic, evidentiary, procedural, and institutional dimensions. The paper defines "delegation envelopes" as pre-authorized regions of action space and "misclosure" as a distinct failure mode where a plausible output cannot be ratified due to underspecified or violated contracts. It also outlines benchmark metrics to test when closure interventions are more effective than additional inference-time search, using a travel rebooking example to illustrate the four-contract stack.
Key takeaway
For research scientists developing AI agents for institutional deployment, you should prioritize designing systems that explicitly compile human intent into inspectable contracts. Focusing on reducing "closure gaps" across semantic, evidentiary, procedural, and institutional dimensions will likely improve time-to-authorized-action and compliance more cost-effectively than merely increasing inference-time search, especially for high-risk tasks. Implement a four-contract stack to define clear delegation envelopes, ensuring auditable and safe autonomous operation.
Key insights
Intent compilation transforms human purpose into binding artifacts for open-world AI agents, addressing deployment failures.
Principles
- Open-world agents require explicit authorization.
- Closure gaps define unresolved task conditions.
- Delegation envelopes enable auditable autonomy.
Method
Intent compilation produces a contract tuple $(S_t, E_t, M_t, I_t)$ specifying task semantics, evidence, method, and authority, exposing residual closure gaps for autonomous action, clarification, or escalation.
In practice
- Define explicit semantic contracts for AI tasks.
- Implement evidentiary rules for data sources.
- Specify procedural contracts for tool use.
Topics
- Agentic AI
- Intent Compilation
- Closure Gaps
- Delegation Envelopes
- AI Governance
Best for: Research Scientist, AI Scientist, AI Architect, MLOps Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.