vercel/ai · error · Error

Deep Agents reasoning requires ChatAnthropic

Error message

Deep Agents reasoning requires ChatAnthropic

What it means

The deepagents bridge's model middleware only supports Anthropic models. When a turn carries thinking/effort settings (but no explicit model string), the bridge resolves the request's model and requires it to be a LangChain `ChatAnthropic` instance; anything else (e.g. an OpenAI or other LangChain chat model) is rejected. Deep Agents reasoning here is wired exclusively through the Anthropic client, so other providers cannot be used for reasoning.

Source

Thrown at packages/harness-deepagents/src/bridge/index.ts:116

        const configuredModel = buildModel({
          rawModel: activeModel,
          thinking: activeThinking,
          effort: activeEffort,
        });
        if (!configuredModel) throw new Error('Deep Agents model is missing');
        return handler({ ...request, model: configuredModel });
      }

      let model = request.model;
      if (
        '_getModelInstance' in model &&
        typeof model._getModelInstance === 'function'
      ) {
        model = await model._getModelInstance();
      }

      if (!(model instanceof ChatAnthropic)) {
        throw new Error('Deep Agents reasoning requires ChatAnthropic');
      }

      const configuredModel = buildModel({
        rawModel: model.model,
        thinking: activeThinking,
        effort: activeEffort,
      });
      if (!configuredModel) throw new Error('Deep Agents model is missing');

      return handler({ ...request, model: configuredModel });
    },
  });
}

const args = parseArgs(argv.slice(2));
const workdir = args.workdir;
const bridgeStateDir = args.bridgeStateDir;
if (!workdir || !bridgeStateDir) {

View on GitHub (pinned to 69428b1f8b)

Solutions

1. Use a `ChatAnthropic` model (from `@langchain/anthropic`) for the Deep Agents agent when thinking/effort is enabled.
2. Set the model explicitly via the harness `model` option (a model slug string) so the bridge builds a `ChatAnthropic` itself via `buildModel`.
3. Remove `thinking`/`effort` from the request if you must use a non-Anthropic model, so the middleware passes through without the instanceof check.
4. Point `ANTHROPIC_BASE_URL` at a gateway if you need a non-Anthropic backend behind an Anthropic-compatible API.

Example fix

// before
new ChatOpenAI({ model: 'gpt-4o' }) // with thinking enabled
// after
import { ChatAnthropic } from '@langchain/anthropic';
new ChatAnthropic({ model: 'claude-sonnet-4-5', thinking: { type: 'enabled', budget_tokens: 1024 } })

import { ChatAnthropic } from '@langchain/anthropic';
if (thinkingEnabled && !(resolvedModel instanceof ChatAnthropic)) {
  throw new Error('Deep Agents reasoning requires ChatAnthropic');
}

function isChatAnthropic(m: unknown): m is ChatAnthropic {
  return m instanceof ChatAnthropic;
}

try {
  await runTurn(start);
} catch (error) {
  if (error instanceof Error && error.message.includes('requires ChatAnthropic')) {
    // switch model to ChatAnthropic or drop thinking/effort and retry
  }
  throw error;
}

When it happens

Trigger: Calling a harness turn with `thinking` or `effort` set while the resolved request model is not a `ChatAnthropic` instance (e.g. `ChatOpenAI`, or a lazy model whose `_getModelInstance()` yields a non-Anthropic client).

Common situations: Configuring the harness/agent with a non-Anthropic LangChain chat model while enabling extended thinking or effort; switching providers after the model middleware was added; passing a wrapped model object that delegates to a non-Anthropic client.

Related errors

• AWS SigV4 authentication requires AWS credentials. Please pr
• AWS SigV4 authentication requires both AWS_ACCESS_KEY_ID and
• ${error.message}
• Anthropic Message Batches do not support per-request betas (

AI-assisted analysis of vercel/ai@69428b1f8b (2026-08-30). Data as JSON: /api/errors/0da5d93d25658b6d. Report an issue: GitHub.