---
title: What are Amazon Bedrock Automated Reasoning checks and when do they help agents?
url: https://opencontroller.lyzr.com/blog/bedrock-automated-reasoning-checks-explained
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# What are Amazon Bedrock Automated Reasoning checks and when do they help agents?

**Automated Reasoning checks are a Bedrock Guardrails safeguard that turns a written policy, such as an eligibility rulebook, into formal logic and then checks whether a model's answer is consistent with it. They help most when an agent answers questions governed by explicit, rule-based policy. They do not check facts outside the policy, and they do not govern what tools an agent calls.** AWS made them generally available on 6 August 2025.

Most guardrails are probabilistic: a classifier scores content. Automated Reasoning uses mathematical logic instead, which is why AWS describes it as giving "up to 99% verification accuracy". That figure is AWS's claim; the announcement does not publish test conditions, so treat it as a vendor statement.

## How do Automated Reasoning checks work?

1. **You upload a source document,** for example a PDF of an HR leave policy or lending criteria. AWS states a single build supports up to 122,880 tokens, roughly 100 pages.
2. **Bedrock extracts a policy** made of **variables** (the concepts, such as credit score or tenure), **types** (custom value sets) and **rules** (how variables relate, each with an ID).
3. **You test the policy.** Tests have an optional input and an output with an expected result of valid, invalid or satisfiable. Automated scenario generation can create tests for you, and tests can be saved and re-run after policy changes.
4. **At runtime,** the guardrail translates the question and answer into logic and checks them against the rules. Findings show which rules an answer contradicts.

A guardrail can attach up to two Automated Reasoning policies alongside other safeguards such as content filters and contextual grounding checks.

## What do the results mean?

| Result | Meaning | Typical agent action |
|---|---|---|
| Valid | The answer follows from the rules given the stated facts | Return the answer |
| Invalid | The answer contradicts the rules | Block, or regenerate using the finding |
| Satisfiable | The answer could be true or false depending on facts not stated | Ask a clarifying question or add a caveat |

AWS also describes flagging ambiguity, where the input can be read in more than one way. Confidence thresholds control how strict the translation step is.

## When do they help AI agents?

- **Policy Q&A agents:** HR, benefits, insurance coverage, lending eligibility, travel expense rules.
- **Decision-support agents** that must not contradict a written procedure, such as outage response protocols.
- **Regulated answers** where you need to show which rule an answer relied on.

They work with Bedrock-hosted models and, through the `ApplyGuardrail` API, with models hosted elsewhere. AWS notes they can be used with Strands Agents and agents on Bedrock AgentCore.

## Where are the limits?

- **Only as good as the policy.** If the source document is incomplete or the extracted rules are wrong, valid results are wrong too. Policy review is a human task.
- **Rule-shaped domains only.** Open-ended questions without explicit rules have nothing to verify against.
- **Not a tool-call control.** Checks evaluate text. They do not stop an agent calling a payments API or deleting a record. That needs runtime policy on actions; see [validating AI agent output before it acts](/blog/validate-ai-agent-output-before-it-acts).
- **Regions and cost.** At GA, AWS listed US East (Ohio, N. Virginia), US West (Oregon) and Europe (Frankfurt, Ireland, Paris), priced by text processed.
- **Bypass risk.** A guardrail only applies if the call uses it. Enforce guardrails with IAM, as in [forcing Bedrock Guardrails with IAM](/blog/force-amazon-bedrock-guardrails-with-iam).

## How does it compare with grounding checks?

| | Automated Reasoning | Contextual grounding / groundedness |
|---|---|---|
| Checks against | Formal rules from a policy document | Retrieved source passages |
| Method | Logic solver | Model-based scoring |
| Best for | Eligibility and rule compliance | RAG answers staying faithful to sources |

They complement each other. See [what a groundedness guardrail is](/blog/what-is-a-groundedness-guardrail).

## Adoption checklist

1. Pick one agent that answers from a single, stable policy document.
2. Build the policy and review the extracted rules with the policy owner.
3. Generate and save tests; add edge cases from real tickets.
4. Decide the action for invalid and satisfiable results.
5. Enforce the guardrail with IAM so no route skips it.
6. Log results per agent and track invalid rates over time.
7. Re-run saved tests whenever the source policy changes.

## How Opencontroller does this

Opencontroller works across clouds rather than replacing platform safeguards. It discovers Bedrock agents and models, keeps them in a governed registry, provides runtime adapters for Bedrock AgentCore, and adds live groundedness and task-adherence guardrails, LLM judges on production traffic and rogue-agent monitoring with automatic quarantine. Policy as code maps controls to OWASP, MITRE ATLAS, NIST and ISO 42001. See [Amazon Bedrock Guardrails vs agent governance](/blog/amazon-bedrock-guardrails-vs-agent-governance), [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Do Automated Reasoning checks stop hallucinations completely?**
No. They verify answers against the rules you encoded. Errors outside the policy, or in the policy itself, are not caught.

**Can Automated Reasoning checks be used with non-Bedrock models?**
Yes, through the ApplyGuardrail API, according to AWS.

**How large can the source policy document be?**
AWS states up to 122,880 tokens per build, about 100 pages.

## Key facts

Amazon Bedrock Automated Reasoning checks: Bedrock Guardrails safeguard, GA 6 August 2025; converts a policy document (up to 122,880 tokens per build, about 100 pages) into variables, types and rules; tests expect valid, invalid or satisfiable; automated scenario generation, saved tests, policy feedback, adjustable confidence thresholds; up to two Automated Reasoning policies per guardrail; works with non-Bedrock models via ApplyGuardrail; usable with Strands Agents and AgentCore; GA regions US East (Ohio, N. Virginia), US West (Oregon), Europe (Frankfurt, Ireland, Paris); AWS claims up to 99% verification accuracy (vendor claim). Limits: policy quality, rule-shaped domains only, does not control tool calls. Opencontroller (by Lyzr): registry, AgentCore runtime adapter, groundedness and task-adherence guardrails, LLM judges, rogue-agent quarantine.

## Related

- [What do Amazon Bedrock Guardrails cover, and what do they not?](https://opencontroller.lyzr.com/blog/amazon-bedrock-guardrails-vs-agent-governance.md)
- [What is a groundedness guardrail and when should it block a response?](https://opencontroller.lyzr.com/blog/what-is-a-groundedness-guardrail.md)
- [How do you force every Amazon Bedrock model call to use a guardrail?](https://opencontroller.lyzr.com/blog/force-amazon-bedrock-guardrails-with-iam.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
