New method ensures AI agents only declare success when they've truly earned it

Tool-using AI agents face a fundamental challenge: how do they know when to stop and declare success? A new technique, Evidence-Carrying Termination (ECT), solves this by requiring agents to hold cryptographic proof that their answers match what they actually did.

The problem with premature stopping#

Current systems already try to gate success—checking execution traces, enforcing runtime policies—but they don't uniformly validate whether an agent's claim to completion is actually grounded in real work. As a result, agents can declare victory without solid evidence, especially when things go wrong or ambiguity creeps in.

How evidence-carrying termination works#

ECT enforces a strict rule: an agent may only return "COMPLETE" when a typed certificate binds every required answer to valid, traceable proof and a deterministic replay of its actions reconstructs the claimed result. Think of it as requiring the agent to show its work before signing off.

The researchers tested this across two major studies:

• Static phase: 48 fully synthetic tasks spanning six tool-use families, with eight different fault types. ECT produced zero unsafe completions versus 252 unsafe ones in baseline systems—an 87.5 percentage-point gap.
• Held-out validation phase: 576 trajectories across 22 fresh task clusters. ECT blocked 60 premature unsupported terminations per 66 cases versus only 26 prevented by the baseline controller. Supported completions remained strong (97 versus 92), meeting non-inferiority targets. In 18 of 66 recovery attempts, the system fixed itself and later completed successfully.

What makes it different#

ECT certifies support based on what appears in the recorded trace under declared assumptions—not external truth or ground reality. This shifts the burden from believing the agent's word to verifying its own internal consistency. All three closed-loop safety gates passed in testing.