ProposeThenValidateAgentGate
title: Propose-Then-Validate Agent Gate type: technique created: 2026-09-20 last_updated: 2026-09-20 related: ["Deterministic-Tool / Agent-Reasoning Partition", "Deterministic Coordination over Agent Adjudication"] sources: ["https://aws.amazon.com/blogs/compute/validating-multi-agent-decisions-with-step-functions-and-bedrock-agentcore/"] radar_quadrant: Techniques radar_ring: Assess radar_position: inner
When an AI agent is allowed to decide things at scale, the risk is that it proposes something that sounds plausible but is invalid: a flight that does not exist, or compensation that breaks a fare rule. The Propose-Then-Validate Agent Gate is a workflow pattern that keeps the agent's judgment but removes its ability to act on it directly. The agent suggests, ordinary code checks the suggestion against real systems and rules, and only checked suggestions move forward.
The AWS Compute Blog example applies the pattern to airline disruption handling, where a cancelled flight leaves hundreds of passengers to rebook. One agent proposes rebooking alternatives, and deterministic code confirms that the proposed flights are actually bookable. A second agent drafts compensation language, and deterministic code checks it against entitlement rules. Edge cases go to a human reviewer through a callback that consumes no compute while it waits. Only deterministic steps perform the final bookings and payments. The post states the principle as "agents propose, and deterministic code validates."
The outcome is that the agent's speed and flexibility are used where they help (drafting options, wording), while the steps with financial or regulatory consequences stay predictable and auditable. The workflow engine records each gate, so a reviewer can see which proposals were rejected and why.
The example is tied to one vendor stack. It notes a single 15-minute limit per agent task and that agent calls are request-response only, with no asynchronous waiting on an agent step. The pattern itself does not depend on those constraints, but the concrete implementation does.
The pattern is placed in Assess. It is a clear, reusable idea and closely related to existing partition patterns on this radar, but the source is a single vendor blog post and there is no first-person production use. It is worth exploring wherever an agent's output triggers an irreversible action.
