One workspace in front of four model providers

Model quality moves quickly, and it does not move in one direction. The best model for long documents this quarter may not be the best for code next quarter. Teams that pick one provider and build around its SDK end up rewriting the same chat twice a year.

AgentJam treats the provider as a setting. The chat, the history, the tools, and the attachments are the same regardless of which model answers. This post explains how that boundary is drawn and what actually changes when you cross it.

A single message shape

Every provider gets the same input: a list of messages with roles, text parts, file parts, and image parts. Every provider returns the same output: a stream of text deltas, optional reasoning deltas, tool calls, and annotations. The adapter for each provider translates in both directions.

const providers = {
  openai: createOpenAI({ apiKey: env.OPENAI_API_KEY }),
  anthropic: createAnthropic({ apiKey: env.ANTHROPIC_API_KEY }),
  google: createGoogleGenerativeAI({ apiKey: env.GOOGLE_API_KEY }),
  xai: createXai({ apiKey: env.XAI_API_KEY }),
} as const;

export const modelFor = (provider: keyof typeof providers, id: string) =>
  providers[provider](id);

Because the shape is shared, the front end has no idea which provider is answering. It renders deltas and annotations. That is the whole contract.

What stays the same

• Streaming. Reasoning and output events look identical regardless of source.
• Attachments. File and image parts are encoded once and passed through the adapter.
• Tools. Web Search and Deep Research are defined once and offered to whichever model is active.
• History. Conversations are stored as the shared message shape, so switching models mid-thread keeps the full context.

What changes

Honesty matters here, because “any model” is easy to say and only mostly true.

Reasoning detail. Some models expose rich planning notes; others expose a summary or nothing. The reasoning block adapts: it shows what the model provides and stays hidden when there is nothing to show.

Context window. Very large attachments may fit one model’s context window and not another’s, so the practical attachment limit depends on the provider.

Tool behavior. All four providers support tool calls, but they differ in how eagerly they use them. The tool picker pins a tool to the message so the choice is explicit rather than left to the model’s mood.

Speed and price. These vary more than quality does. Per-deployment configuration lets a team standardize on one provider for cost, or route by task.

Choosing a default

For a new deployment, the recommendation is boring: pick the provider your organization already has an agreement with, set it as the default, and revisit quarterly. The point of the abstraction is not to switch constantly. It is to make switching a configuration change instead of a rewrite when you do want to.

Try it

Ask the same question in two workspaces backed by different providers and compare the reasoning blocks. The answers will often agree. How they got there is where the differences live.