IBM JevAct Adds Decision Gates to Agents

JevAct is a compelling harness-level reliability upgrade because it leaves the agent model, prompt, and tools unchanged. The results are promising, but this is a preflight layer—not a substitute for hard authorization and deterministic security controls.

• –An offline LLM compiler turns domain policies into executable checks, decision rules, questions, and hand-off checklists.
• –The gate can block or revise risky writes before they commit, catching policy violations and premature agent abandonment.
• –Telecom performance improved sharply, while Airline and BFCL results were more mixed.
• –Ablations suggest programmatic checks contributed more than probabilistic questions in some settings.
• –Added judge calls, policy-compilation errors, threshold calibration, and latency remain important deployment risks.