---
title: Why can an agent not invoke a Bedrock model you can see in the console?
url: https://opencontroller.lyzr.com/blog/bedrock-model-access-and-inference-profiles-for-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# Why can an agent not invoke a Bedrock model you can see in the console?

**Seeing a model in Amazon Bedrock does not mean an agent can call it. The usual blockers are an Anthropic first-use form that was never submitted, a model that must be invoked through an inference profile ID rather than its model ID, IAM policies that allow the profile but not the foundation model in every destination Region, and SCPs whose Region rules block cross-Region or global routing.**

This is one of the most common problems when routing agents or a gateway through Bedrock. The model list looks complete, yet requests fail with access denied or validation errors. Each cause has a specific fix. This guide is part of our [guide to governing AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms).

## Is model access still something you have to enable?

Mostly no. In October 2025 AWS changed Bedrock so that serverless foundation models are enabled by default in all commercial Regions where Bedrock is supported. You no longer request access model by model.

There is one important exception. Anthropic models are enabled but need a one-time usage form before first use. It can be submitted through the API or by choosing an Anthropic model in the console playground. If the organisation's management account submits it, Anthropic models are enabled across member accounts. If an agent in a new account fails on its first Claude call, check this first.

Because access is now on by default, restriction moves to IAM and SCPs. That is good for governance, but it means the policies themselves are usually what blocks an agent.

## Why does the model ID fail when the inference profile works?

Some models in some Regions cannot be called on demand by their plain model ID. The error says on-demand throughput is not supported and asks for the ID or ARN of an inference profile. Inference profiles route a request to one of several Regions:

| Profile type | ID prefix example | Where requests can run |
|---|---|---|
| Geographic cross-Region | `us.`, `eu.`, `jp.`, `au.` | Regions within that geography |
| Global cross-Region | `global.` | Supported commercial Regions worldwide |
| Application inference profile | Your own ARN | Wraps a model or system profile, often for cost tags |

Agents, SDK wrappers and gateways often store the bare model ID. Update the configuration to the inference profile ID for those models, and choose geographic or global deliberately. Global profiles can route outside your geography, which matters for data residency.

## Which IAM permissions does cross-Region inference need?

This is where most failures come from. According to AWS guidance on securing cross-Region inference:

1. **Allow the inference profile** in the source Region, for example `bedrock:InvokeModel` on the `us.` profile ARN.
2. **Allow the foundation model in every destination Region** the profile can route to, not just the source Region. Scope this with the `bedrock:InferenceProfileArn` condition so the model can only be reached through that profile.
3. **For global profiles,** also allow the Region-less foundation model ARN (`arn:aws:bedrock:::foundation-model/...`) with `aws:RequestedRegion` set to `unspecified` and the `bedrock:InferenceProfileArn` condition.

A policy that allows only the model in us-east-1 will fail intermittently when the profile routes a request to us-east-2 or us-west-2. Intermittent failures are a strong hint that this is the cause.

## How do SCPs break cross-Region inference?

Many landing zones use SCPs that deny actions outside approved Regions using `aws:RequestedRegion`. For a geographic profile, the SCP must allow every destination Region in it. For global profiles, requests carry `unspecified` as the requested Region, so a Region deny list blocks them unless Bedrock is handled separately.

AWS's suggested pattern is to exclude Bedrock from the broad Region-deny statement and add a Bedrock-specific statement. You can also use `bedrock:InferenceProfileArn` with `ArnLike` patterns to block whole profile families, for example denying all `global.*` profiles where residency rules require it. If you use AWS Control Tower, make these changes through Customizations for AWS Control Tower rather than editing managed SCPs, which causes drift. Our guide to [restricting Bedrock models with SCPs](/blog/restrict-amazon-bedrock-models-with-scps) has more examples.

## How do you see where a request actually ran?

CloudTrail logs cross-Region calls in the source Region. The event's `additionalEventData` includes `inferenceRegion`, which shows the Region that served the request. Use it to prove residency and to debug failures tied to one destination Region. See also [auditing Amazon Bedrock model calls](/blog/audit-amazon-bedrock-model-calls).

## What checklist should you run when an agent call fails?

1. Is it an Anthropic model in an account that has not submitted the first-use form?
2. Is the agent using a model ID where the model needs an inference profile ID?
3. Does IAM allow the profile and the foundation model in every destination Region?
4. Does an SCP Region rule block a destination Region or `unspecified`?
5. Is the request format right for the API used (InvokeModel body vs Converse)?
6. Does CloudTrail show the denied action and Region?

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and offers a runtime adapter for Bedrock AgentCore. Its LLM Gateway (Lyzr Gateway) routes model traffic with fallback, retry, caching and cost attribution per agent and per version. Posture flags risky permission and configuration patterns. See [routing Bedrock models through an LLM gateway](/blog/route-bedrock-models-through-an-llm-gateway), the [capability reference](/capability-reference), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Do I still need to request access to Bedrock models?**
Not for most serverless models since October 2025. Anthropic models still need a one-time usage form before first use.

**Why do Bedrock calls fail only some of the time?**
Often because a cross-Region inference profile routes some requests to a Region where IAM or an SCP does not allow the foundation model.

**Can I stop agents using global inference profiles?**
Yes. Deny global profiles with an SCP using `bedrock:InferenceProfileArn` matching `global.*`, or simply do not grant global permissions in IAM.

## Key facts

- Since October 2025, Amazon Bedrock serverless foundation models are enabled by default in commercial Regions; Anthropic models still require a one-time usage form.
- Some models must be invoked through an inference profile ID (geographic like us./eu., or global.) rather than the plain model ID.
- Cross-Region inference needs IAM permission on the inference profile and on the foundation model in every destination Region, scoped with bedrock:InferenceProfileArn.
- Global inference requests carry aws:RequestedRegion = unspecified, so Region-deny SCPs block them unless Bedrock is handled separately.
- CloudTrail logs cross-Region calls in the source Region with inferenceRegion in additionalEventData.
- Opencontroller auto-discovers models and keys across AWS, Azure and Google Cloud and has a Bedrock AgentCore runtime adapter.
- Opencontroller starts from $250K per deployment per year.

## Related

- [Why do Amazon Bedrock calls fail through an LLM gateway, and how do you fix them?](https://opencontroller.lyzr.com/blog/route-bedrock-models-through-an-llm-gateway.md)
- [How do you restrict which Amazon Bedrock models agents can use?](https://opencontroller.lyzr.com/blog/restrict-amazon-bedrock-models-with-scps.md)
- [How do you audit Amazon Bedrock model calls made by AI agents?](https://opencontroller.lyzr.com/blog/audit-amazon-bedrock-model-calls.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
