LACUNA: Safe Agents as Recursive Program Holes

Researchers introduced LACUNA, a safe programming model for LLM agents that lets the model fill typed code holes at runtime while enforcing checks before execution.

Unsafe generations are rejected before any side effects, and the framework can express ReAct loops, sub-agents, skills, planning, and parallel decomposition.

In benchmarks, LACUNA sometimes retried a small number of actions, blocked risky outputs early, and performed close to baseline on task suites like BrowseComp-Plus and tau^2-bench.

The result is a more expressive agent design that preserves controlled tool access and safer runtime behavior without giving up much performance.