AWS adds strict JSON and tool schemas for Claude on Bedrock
AWS adds structured Claude outputs and strict tool schemas to Bedrock. Why this matters for agents, plus a documented Mantle endpoint discrepancy.
On October 19, AWS announced structured outputs for Anthropic Claude on Amazon Bedrock. JSON Schema and Strict Tool Use constrain response structure and tool-call inputs during generation. They can reduce formatting failures in agent workflows, but do not establish factual accuracy or authorization. AWS's blog and user guide currently give conflicting guidance for the Bedrock Mantle endpoint.
Prompt-only formatting often needed parsing, retries and validation when model outputs contained missing fields, unexpected data types or malformed tool parameters.
Supported Bedrock APIs can constrain generated data and function inputs to a declared JSON schema. Endpoint and model support still need explicit verification.
Why agent developers should care about schema fidelity
An AI agent does more than write paragraphs. It submits service tickets, updates business records and passes arguments to external software. A function expecting an integer can reject a natural-language phrase even when a human understands the intended value. AWS now offers a generation-time constraint: it validates a supported schema and compiles an output grammar so normal decoding follows the requested field types and structure. This can eliminate a class of integration failures that used to surface after the model returned a response.
Two capabilities solve different problems. JSON Schema output format constrains the shape of the final answer. Strict Tool Use constrains the parameters of tool calls. They can be used together, but neither proves that the model extracted the correct account number or that a requested operation is authorized. A schema-valid request can still be dangerous or factually wrong. Runtime permissions, business validation and audit logs remain necessary layers.
Caching, unsupported constraints and failure cases
AWS validates the submitted schema and compiles a grammar the first time a new structure is used. Compilation may increase first-call latency; the compiled result is then cached for 24 hours within an account. Reusing stable schemas is more efficient than changing field structure continuously. AWS also requires `additionalProperties: false` on object definitions.
The important limitations are easy to miss. The documentation says numerical boundaries such as `minimum` or `maximum` and string length limits such as `minLength` are not enforced during generation. The AWS blog also advises handling refusals and responses cut short by maximum-token limits. Structured outputs cannot be combined with citations for Anthropic models according to the user guide. No team should treat schema fidelity as proof that a generated statement is complete, supported by evidence or safe to execute.
A real documentation conflict on the Mantle endpoint
There is a notable mismatch between two first-party AWS sources. The October 19 blog post describes structured Claude output and Strict Tool Use on both `bedrock-runtime` and `bedrock-mantle`. The Amazon Bedrock user guide, however, states that `output_config.format` on the Anthropic Messages API at `bedrock-mantle` returns HTTP 400. Without an independent endpoint test, we cannot say which description reflects the current behavior.
A prudent implementation would first confirm structured outputs through supported Converse or InvokeModel paths on `bedrock-runtime`, check model and Region support, and test Mantle separately before declaring compatibility. A meaningful pilot should include malformed data types, missing fields, model refusals, output truncation and attempted unauthorized actions. That distinguishes a real reliability gain from the mistaken assumption that valid JSON means an agent can safely act.
AgentenCode assessment
AWS is making agent interfaces more predictable by enforcing structure during generation. The remaining hard part is not formatting: it is whether the proposed action is true, appropriate and authorized. The unresolved Mantle documentation mismatch deserves direct testing rather than optimistic feature claims.
Sources and verification
Based on AWS's October 19, 2026 technical announcement and the Amazon Bedrock user guide, which currently differ regarding Mantle. AgentenCode has not run an independent AWS account test.