New Framework Blocks AI Agents From Acting on Incomplete Instructions

Autonomous agents often fail when following vague or incomplete human instructions, even after identifying what information is missing. Researchers have discovered that detection alone doesn't solve the problem—agents frequently proceed with flawed actions anyway.

The Detection-Execution Gap#

Current approaches to improving agent reliability focus on teaching models to recognize incomplete or ambiguous instructions. However, this detection remains disconnected from execution behavior: an agent might identify that critical details are missing, yet still attempt to carry out the task with guesses or assumptions. This gap between knowing something is wrong and actually preventing action has led to hallucinated actions and unsafe outcomes in real deployments.

How Cocoreli Works#

A new framework called Cocoreli restructures this problem by architecturally coupling detection and prevention. Rather than relying on model judgment, Cocoreli blocks execution by design when required parameters remain unresolved. The system models task structure, tracks what information is missing, and refuses to proceed until gaps are filled through targeted clarification dialogue.

The key innovation is structural: detecting a missing parameter simultaneously prevents action. This enforced precondition model removes the behavioral gap that bedevils reasoning-based approaches.

Testing Against Alternatives#

Evaluation in a controlled construction environment demonstrates the difference. Chain-of-thought reasoning, prompt-chaining, and ReAct-style agents often execute despite incomplete specifications, even when detection rates appear high. Cocoreli eliminates hallucinated actions entirely by blocking execution until specifications resolve. The same modular approach also generalizes to API workflow tasks on benchmark datasets like ToolBench.

Implications#

The research suggests that reliable collaborative execution between humans and agents requires more than improving model capability—it demands architectural enforcement of constraints. Safety and correctness may ultimately rest not on what models can reason about, but on what systems refuse to do without proper authorization.