Choose a model and reasoning

Strands harness runs on the latest models across Amazon Bedrock, Anthropic, OpenAI, and Google, and uses Amazon Bedrock by default. Pass model to pick another; pass effort to change how hard the agent reasons. Both are single arguments that work the same across every provider.

Point Strands harness at a model

You pick a model by passing model, which accepts three forms: a provider/name string, a bare Bedrock model id, or a ready-made Strands Model instance. The provider/ prefix has aliases for bedrock, bedrock-mantle, anthropic, openai, google, ollama, and litellm. A string with no provider prefix is treated as a bare Amazon Bedrock id.

Strands harness

Python

from strands_harness import create_harness

agent = create_harness(model="openai/gpt-5.6-sol")

TypeScript

import { createHarness } from '@strands-agents/harness'

const agent = await createHarness({ model: 'openai/gpt-5.6-sol' })

SDK

Python

from strands import Agent

agent = Agent(model="global.anthropic.claude-sonnet-5")

TypeScript

import { Agent } from '@strands-agents/sdk'

const agent = new Agent({ model: 'global.anthropic.claude-sonnet-5' })

For a provider Strands harness does not have an alias for, or for a model you have already configured (custom endpoint, credentials, request fields), build the Model instance yourself and pass it. Strands harness uses it as-is:

Python

from strands.models import BedrockModel
from strands_harness import create_harness

model = BedrockModel(model_id="global.anthropic.claude-opus-4-8", region_name="us-west-2")
agent = create_harness(model=model)

TypeScript

import { BedrockModel } from '@strands-agents/sdk'
import { createHarness } from '@strands-agents/harness'

const model = new BedrockModel({ modelId: 'global.anthropic.claude-opus-4-8' })
const agent = await createHarness({ model })

An unknown provider prefix fails at construction with the list of supported providers, so a typo surfaces immediately rather than as a request error later.

Set reasoning effort

effort maps one reasoning level onto whatever each provider’s API expects, so you set it once regardless of provider:

• "auto" (the default) uses each provider’s recommended level.
• "low", "medium", and "high" set it explicitly.
• "off" (or off on the CLI) turns reasoning off.

Strands harness

Python

from strands_harness import create_harness

agent = create_harness(model="anthropic/claude-opus-4-8", effort="high")

TypeScript

import { createHarness } from '@strands-agents/harness'

const agent = await createHarness({ model: 'anthropic/claude-opus-4-8', effort: 'high' })

SDK

Python

from strands import Agent
from strands.models import BedrockModel

bedrock_model = BedrockModel(
    model_id="global.anthropic.claude-sonnet-5",
    additional_request_fields={
        "thinking": {"type": "enabled", "budget_tokens": 4096}
    },
)
agent = Agent(model=bedrock_model)

TypeScript

import { Agent, BedrockModel } from '@strands-agents/sdk'

const bedrockModel = new BedrockModel({
  modelId: 'global.anthropic.claude-sonnet-5',
  additionalRequestFields: {
    thinking: { type: 'enabled', budget_tokens: 4096 },
  },
})
const agent = new Agent({ model: bedrockModel })

A level a provider does not accept fails at construction rather than mid-request. Some providers accept finer levels (for example minimal or xhigh); Strands harness validates against the resolved provider’s own set, so use the levels that provider documents.

Reasoning does not apply to a pre-built Model instance: configure reasoning on the instance itself when you pass one.

Note on web search and caching

Two model-level features follow from the provider you pick. Native web_search works on OpenAI, Anthropic, Google, and GPT-5/GPT-6 models on bedrock-mantle; prompt caching is configured by Strands harness on Amazon Bedrock and Anthropic direct and is automatic elsewhere. Both are covered where you enable them: web access and manage context and caching.

For the full option list, see the configuration reference.