Is this the same as deploying OpenAI?

No. Deploying OpenAI is the model and residency decision, including Azure OpenAI. This is the agent runtime on top of it: schemas, tools, and where the decision gets validated before anything is written.

Do we need an SDK at all?

Not for a single deterministic call. You need one when the step has tools, retries, and a shape another system must accept.

Will you lock us to OpenAI?

No. We route legal and long-context work elsewhere routinely. The gateway is what makes that a configuration instead of a rewrite.

AI stack

Structured output is a contract, not a hope.

The first week is in the org, reading what already exists. The OpenAI Agents SDK is strongest where the step has to produce a shape another system will accept: a routing decision, an extracted record, a tool call with typed arguments. The honest no: if the work is a software factory with a named delivery date, we are the wrong partner.

Send the brief

What we put in

Typed arguments, then a boring handler.

01
Define the shape
A JSON schema for the decision. If the output shape cannot be written down, the step is not ready for an agent.
02
Let the model fill it
Function calling and structured outputs, so parsing prose out of a paragraph stops being part of the system.
03
Validate, then act
Rejected output is an exception with an owner, not a best-effort write.
04
Keep handlers dull
The handler is ordinary code: testable, idempotent, and reviewable. That is the point.

Structured routing with the Agents SDK. The model produces a typed decision against a JSON schema, a deterministic handler executes it, and anything the schema rejects goes to a person instead of into a record.

Mermaid source

flowchart TD
  A["Request: message, document, or event"] --> B["Agent with a JSON schema"]
  B --> C["Typed decision and arguments"]
  C --> D["Schema validation"]
  D -->|"rejected"| E["Exception queue: a person decides"]
  D -->|"accepted"| F["Deterministic handler or tool call"]
  F --> G["Write through the durable plane"]

When it earns the seat

Routing, extraction, and high-volume classification.

Governed

The same controls as every other runtime.

Questions

What we actually say.

Is this the same as deploying OpenAI?: No. Deploying OpenAI is the model and residency decision, including Azure OpenAI. This is the agent runtime on top of it: schemas, tools, and where the decision gets validated before anything is written.
Do we need an SDK at all?: Not for a single deterministic call. You need one when the step has tools, retries, and a shape another system must accept.
Will you lock us to OpenAI?: No. We route legal and long-context work elsewhere routinely. The gateway is what makes that a configuration instead of a rewrite.

What to read next.

The brief

Send the shape the next system will accept.

We will tell you whether a typed agent step is the cheapest way to produce it.

Structured output is a contract, not a hope.

flowchart TD A["Request: message, document, or event"] --> B["Agent with a JSON schema"] B --> C["Typed decision and arguments"] C --> D["Schema validation"] D -->|"rejected"| E["Exception queue: a person decides"] D -->|"accepted"| F["Deterministic handler or tool call"] F --> G["Write through the durable plane"]