Strands Harness: Frontier Performance with $\approx 28\%$ Lower Token Costs

Published: September 21, 2026 | Read Time: 6 min

Strands harness is a state-of-the-art, fully assembled, and highly customizable agent framework. Whether you prefer to run it locally or deploy it anywhere in the cloud, it provides a seamless experience for building autonomous agents.

---

The Motivation

The development team—including Arron Bailiss, Albert Zhao, Tim Moreton, Gautam Sirdeshmukh, Murat Kaan Meral, and Ariel Nabavian—observed a recurring pain point for developers. Many builders found that their agent concepts "just worked" when using local harnesses like Claude Code or Codex, but they struggled to replicate that fluidity in a cloud environment.

"It’s tricky wiring up the right primitives just well enough to match that 'it just worked' feeling."

To solve this, they created Strands harness. While many existing tools are strictly coding agents, Strands is designed as a general-purpose agent.

---

Performance & Benchmarking

Strands harness is engineered for cost-efficiency without compromising on precision. Across six different benchmarks, it achieved a token cost reduction of $\text{Cost}_{\text{reduction}} = 28\%$ compared to other harnesses using the same GPT or Claude models.

Comparative Analysis

The team utilized distributed benchmarking on EC2 with Harbor to validate these claims.

Notably, when using Fable 5, Strands harness not only cost 77% less than Claude Code but actually outperformed it on Terminal Bench 2.1.

---

Technical Deep Dive: The SDK

Built upon the Strands Harness SDK, the system implements sophisticated defaults for context management and prompt caching.

Context Management Logic

The efficiency is driven by three primary mechanisms:

1. Truncation: Tool results exceeding $\sim 1500$ tokens are automatically truncated.
2. Compaction: Summarization triggers once the context window exceeds $85\%$.
3. Recovery: A context recovery process runs within the loop to handle overflows.

A formal research paper detailing these benchmarks is expected to follow shortly.

---

Getting Started

Integration is designed to be trivial, requiring only a single line of code in either Python or TypeScript.

Supported Providers

The harness is compatible with:

• Amazon Bedrock
• Anthropic
• OpenAI
• Google
• Ollama (Local)
• LiteLLM

Implementation Examples

1. Cloud Providers (Bedrock, Anthropic, OpenAI, Google)

Python

from strands_harness import create_harness

# Works with "bedrock/global.anthropic.claude-opus-5", 
# "anthropic/claude-opus-5", "openai/gpt-5.6-sol", or "google/gemini-3.5-flash"
agent = create_harness(model="bedrock/global.anthropic.claude-opus-5")
agent("Research the top three vector databases, compare pricing and limits, and write it up in comparison.md")

TypeScript

import { createHarness } from "@strands-agents/harness";

const agent = await createHarness({ model: "bedrock/global.anthropic.claude-opus-5" });
await agent("Research the top three vector databases, compare pricing and limits, and write it up in comparison.md");

2. Local Models (Ollama)

Python

from strands_harness import create_harness
from strands.models.ollama import OllamaModel

model = OllamaModel(host="http://localhost:11434", model_id="llama3.1")
agent = create_harness(model=model)
agent("Research the top three vector databases, compare pricing and limits, and write it up in comparison.md")

TypeScript

import { createHarness } from "@strands-agents/harness";

const agent = await createHarness({ model: "ollama/llama3.1" });
await agent("Research the top three vector databases, compare pricing and limits, and write it up in comparison.md");

3. LiteLLM Integration

Python

from strands_harness import create_harness
agent = create_harness(model="litellm/openai/gpt-5.6-sol")
agent("Research the top three vector databases, compare pricing and limits, and write it up in comparison.md")

TypeScript

import { createHarness } from "@strands-agents/harness";
const agent = await createHarness({ model: "litellm/openai/gpt-5.6-sol" });
await agent("Research the top three vector databases, compare pricing and limits, and write it up in comparison.md");

---

Summary of Capabilities

When you call create_harness(), you receive an agent that is ready for production immediately:

• Multi-Provider Support: Compatible with Bedrock, Anthropic, OpenAI, Google, Ollama, and LiteLLM.
• Reasoning Capabilities: Optimized for the latest frontier reasoning models.
• Integrated Toolset:
  • shell access
  • file operations (read, write, and edit)
  • web browsing tools