Adovo

Applied AI

A tool schema can change meaning across an agent model adapter

Verify strict tool schemas at the provider boundary when running the OpenAI Agents SDK through LiteLLM.

November 18, 2026

A schema that looks strict in application code can reach a model provider with different constraints. A reported Agents SDK adapter issue described strict being dropped on a LiteLLM path to Anthropic. That issue is now closed, so it is a historical example, not evidence of a current universal defect. The Agents SDK model guide calls LiteLLM integration best-effort beta and advises testing tool calling and structured outputs with the chosen backend.

Inspect the request at the boundary

Choose one production tool with required fields and a disallowed extra field. Run it through the exact adapter, provider, model, and library versions planned for release. Capture the outbound schema in a safe test environment and compare it with the application's definition. A passing local JSON schema test says little about translation between APIs.

Validate the response again

Send test prompts that omit a required value, invent an extra key, and use an unexpected type. Reject invalid tool arguments in your own runtime before side effects. Provider-side constraints can reduce bad calls, but the application still owns the final contract. Record the rejected shape and adapter version without logging private customer input.

Treat upgrades as contract changes

After a model or adapter upgrade, rerun the same acceptance set. Keep a small version matrix with observed wire schemas and invalid-call behavior. If a provider cannot honor a specific constraint, adjust the tool design or enforce the constraint locally, with that limitation visible to the owner.

Adovo’s view

Reliability lives at the boundary where intent becomes a request. A typed tool definition is a starting point; the provider request and the side effect are what the customer experiences.

Continue exploring

Applied AI

OpenAI Agents SDKAgent Reliability

An agent memory limit can cut a reasoning response in half

Test how OpenAI Agents SDK session limits trim reasoning and message items before long-lived agents reach production.

Read more

Applied AI

LangGraphStreaming

A hidden Location header can make an agent stream look finished

Test LangGraph JS run-stream reconnection through proxies and CORS before trusting a clean close.

Read more

Applied AI

LangGraphAgent Reliability

A LangGraph Postgres checkpoint read can scan old blob versions

Investigate checkpoint read latency that grows with a thread’s history in the JavaScript Postgres saver.

Read more
Explore applied AI services →