vercel-ai adapter ​
Status: shipped, release phase. It passes the contract suite in chat mode, on a fake generateText with a model that plays the Acme example agent, and on the real generateText of the ai package (7.0.137) with its MockLanguageModelV4 playing the same agent. It has not run against a live model yet.
The vercel-ai adapter runs an agent built with the Vercel AI SDK: a model, a system prompt and tools, run by generateText. You hand it the SDK's generateText and your model; ai is an optional peer dependency of @agent-nexus/driftbench-adapters; driftbench does not depend on it.
Options ​
import { generateText } from 'ai';
import { openai } from '@ai-sdk/openai';
import { vercelAiAdapter } from '@agent-nexus/driftbench-adapters';
import { salesPrompt, salesTools } from './agent'; // your prompt, and a function that returns the tools
const adapter = vercelAiAdapter({
generateText,
model: openai('gpt-5'),
system: salesPrompt,
tools: (given) => salesTools(), // { coverage_check: tool({ ... }), ... }
toolLabels: { coverage_check: 'coverage.check', order_create: 'order.create' },
sideEffects: ['order.create'],
});| Option | Meaning |
|---|---|
generateText | the SDK's generateText |
model | the language model |
system | the system prompt. System messages in the history are appended to it. |
tools | the tools by name, or a factory called once per session |
maxSteps | the most steps (model calls) one block may take, through stopWhen. Default 8. A block's budget.steps lowers it. |
toolLabels | the tool names mapped to spec labels; a name the providers accept takes no dot |
retry, sleep | the adapter's own retry settings, and how a wait is spent (tests) |
sideEffects, deterministic, version | as on the openai-agents adapter |
What the agent receives ​
Each block calls generateText once with the session history as model messages: a user message as text, a recorded tool call as an assistant tool-call part and a tool message with its tool-result ({ type: 'json', value }, or { type: 'error-text', value } for an error), then the reply text. The SDK closes the agent loop, running each tool's execute.
The adapter adds two wrappers, and changes neither of your objects:
- Each tool's
executeanswers a forced result or error from the block'sworld.toolsfirst, by spec label. The wrapper is a new tool object. - The model's
doGenerateruns under the shared retry policy, through a proxy that keeps the model's class and fields;generateTextruns withmaxRetries: 0, so each model call is retried once, by the policy, and recorded.
What the adapter reads ​
The steps of the result, in order: text parts become text, tool-call parts tool calls, tool-result parts their outputs, tool-error parts their errors (the thrown message). Reasoning parts add nothing. Each step's usage.inputTokens and outputTokens sit on its first segment. raw.finishReasons lists each step's finish reason. A call the run never answered gets an error result.
Capabilities ​
seedPrefix and toolStubs yes; endState with sideEffects; variants, trace and sandbox no; deterministic as declared.
Errors ​
| Error | Becomes |
|---|---|
a 408, a 429 or a 5xx (statusCode), or a network error, after the policy's attempts | AdapterTransportError, marked exhausted |
any other error generateText throws | AgentFailure |
Target ​
The adapter reads no field of the target except retry. Its id is vercel-ai.