AWS split AI agents from execution: agents propose, deterministic code decides what happens

AWS has introduced a new architectural approach to agentic workflows that deliberately separates the reasoning work of AI agents from the power to execute changes. The company released a Step Functions pattern and supporting case study this week arguing that as agents take on higher-stakes tasks—like rebooking passengers or handling payments—deterministic code should always validate agent proposals before anything actually happens.

The airline rebooking pattern: agents suggest, code approves#

The clearest example comes from a new Step Functions workflow pattern for airline disruptions. When a flight is cancelled, Amazon Bedrock AgentCore agents can analyze the situation and propose alternative routes, draft compensation language, and suggest next steps. But no reservation gets changed and no payment goes out until deterministic validation steps run and approve the proposal.

AWS describes the principle plainly: "agents propose, and deterministic code validates."

This separation moves orchestration, routing, and validation logic out of the agent layer and into Step Functions workflows. Each agent runs, generates a proposal, and then hands off to deterministic code that audits the decision. The workflow maintains a complete execution history for compliance and review. Because agents can never directly touch reservations or payments, this pattern naturally prevents an agent hallucination or error from resulting in unauthorized changes.

The same approach could extend to other high-stakes domains—accounts payable, procurement, regulatory filings—anywhere deterministic software should be the final gatekeeper between an AI recommendation and a real-world action.

Code as a computational scratchpad for reasoning#

On the same day, AWS published a case study with Abnormal AI, an email security platform, showing another way code and reasoning should team up. Abnormal AI uses Bedrock AgentCore Code Interpreter—a managed, sandboxed runtime—to let agents execute calculations and data transformations during their reasoning process.

The distinction AWS emphasizes is crucial: Code Interpreter isn't just a tool for agents to call; it's a . Large language models excel at semantic reasoning and coherence, but they struggle with precise calculation, data counting, and structured verification. By pairing probabilistic reasoning with deterministic computation, agents can work through tasks that don't map cleanly to language-based logic—converting data into structured reports, validation counts, threat scoring—then verify their answers computationally rather than relying on the model's intuition alone.