Strands Harness: AWS's Open-Source Agent at 28% Less Cost

Strands Harness: AWS's Open-Source Agent at 28% Less Cost

If you have ever prototyped an agent inside Claude Code or Codex and then tried to rebuild it as your own loop, you know the gap. The CLI "just works," and your from-scratch harness quietly does not. On September 21, 2026, the Strands Agents team at AWS released Strands harness to close exactly that gap: a fully assembled, general-purpose agent you start with one line of Python or TypeScript, run locally or deploy to any linux-container provider, under the Apache 2.0 license.

The pitch is a money number: 28% lower token cost than rival harnesses running the same Claude or GPT models, across six benchmarks, at near-equal accuracy. Whether you buy that number or not, the launch is worth reading because it lands in the middle of a real debate about how much the harness — the loop, tools, context handling and recovery around the model — moves your agent bill.

Building your own agent has been the hard way

The Strands org had already shipped the Strands Harness SDK, which exposes the building blocks: the loop, tool registry, context management, memory, and recovery. Harness takes those primitives and packages them into working defaults. It is explicitly built as a general-purpose agent, not a coding agent — the kind of thing you point at a research task, a data-processing job, or an MCP-connected workflow, not just at a repo.

For a team that has hand-rolled an agent loop, that distinction matters. Most "extensible" frameworks give you a few hook points. A packaged harness like this one hands you the whole machine and lets you prune it, which is a different starting point. (If you want the extreme version of this philosophy, the DeepSeek Harness — everything a plugin — is the other end of the spectrum.)

The headline number, with its honest caveat

The team benchmarked on distributed EC2 instances using Harbor, the evaluation framework from the Terminal-Bench creators, across six benchmarks: ALFWorld, ContextBench, GAIA, WebShop, τ²-bench, and Terminal-Bench 2.1. Cost is the average dollars per task; rivals on the chart are Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness.

The 28% figure has a caveat worth knowing before you quote it: DeepSeek Harness was the most token-efficient harness overall, about 14% cheaper than Strands, but it scored lower on every benchmark. Including it in the comparison is exactly why the headline is 28% and not higher. That is the honest way to read a vendor benchmark: the savings claim survives including the cheapest rival, but the rival gives up accuracy to get there.

The cleanest head-to-head: Claude Fable 5 on Terminal-Bench 2.1

MarkTechPost reproduced the team's chart for Claude Fable 5 on Terminal-Bench 2.1, 89 trials per harness:

Against Claude Code, Strands cost 77% less and scored 7.9 points higher. Oh-my-pi hit the same 69.7 accuracy at 54% higher cost. DeepSeek stayed cheaper but trailed by 10.2 points. The genuinely interesting row is Claude Code itself: $248.05 for the same model, a 4x premium over Strands for a lower score.

What actually drives the efficiency

The savings are not magic. Strands ships defaults for prompt caching and context management, and the team says context management largely drove both token-efficiency and accuracy. Three rules do the work:

• Tool results over roughly 1,500 tokens get truncated.
• Summarization (compaction) triggers when context usage passes 85%.
• Context recovery runs inside the loop if the window overflows.

On top of that, the harness offloads bulky tool results to files and caches the reused parts of each request. Long terminal dumps never flood the next model call; the parts that repeat get billed once.

The harness tax is real, and it's bigger than Strands

This is not one vendor's marketing. In September 2026, researchers from UC Berkeley's Sky Lab and Arena published HarnessTax, running seven models through three harnesses (Claude Code, Codex CLI, and the minimal open-source Pi) on SWE-bench Lite and Terminal-Bench 2.0. Their finding: harness choice barely moves task success rate — within ±2% on SWE-bench Lite, about ±5% on Terminal-Bench 2.0 — while the same model can reach similar success at up to 5x the cost.

That is the frame every benchmark claim belongs in. The Strands numbers are consistent with it: same model, different scaffolding, wildly different dollars per task. So when you read "28% cheaper at near-equal accuracy," treat it as testable — and test it on your workloads, because a harness that wins on these six benchmarks may not win on your long-tail tool calls.

Getting started

Strands harness installs with pip install strands-harness or npm install @strands-agents/harness. Pick a model by name, point it at a local Ollama model, or route through LiteLLM:

from strands_harness import create_harness

agent = create_harness(model="anthropic/claude-opus-5")
agent("Research the top three vector databases, compare pricing and limits, and write it up in comparison.md")

Out of the box, create_harness() returns an agent that runs on current reasoning models across Amazon Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM; ships shell, file (read/write/edit) and web tools rather than a bespoke tool per task; manages its own context window; keeps long-term memory across runs and resumes from a session ID; delegates open-ended subtasks to a built-in helper agent while tracking work with a checklist; and loads Agent Skills if it finds them.

If you want to prototype before committing to code, the Strands CLI (npm install -g @strands-agents/cli) lets you describe an agent in plain English. When the behavior is right, /export emits the whole thing as a Python or TypeScript project — the demos show the Playwright MCP server being added mid-conversation and then baked into the exported boilerplate. The CLI itself is built on Strands harness, so the agent you prototype in a chat is the same code that deploys.

For deployment, the harness targets any provider that runs a linux container: Modal, Cloudflare Containers, Azure Container Apps, Google Cloud Run, Amazon ECS, and Bedrock AgentCore.

The takeaway for developers

Bench harnesses, not just models. Strands harness is a well-packaged default and its context-management rules are genuinely good defaults to steal even if you never use Strands. But the larger lesson — the one the HarnessTax study hammers home — is that your agent's cost is decided by your harness, not just by which frontier model you name. Whether Strands' 28% survives contact with your actual workflows is a measurement, not an assumption.

If this space is new to you, our explainers on what an agent harness is and the earlier DeepSeek Harness plugin architecture are good places to orient yourself.

Sources

• Strands Agents: Introducing Strands harness: frontier performance with 28% lower token cost
• MarkTechPost: AWS Strands Agents Team Releases Strands Harness: An Open-Source Agent Harness With 28% Lower Token Cost at Comparable Accuracy
• Strands Agents: strands-agents/harness-sdk (GitHub)
• Strands Agents: Strands harness docs
• UC Berkeley Sky Lab / Arena: HarnessTax: How Much Does the Harness Matter for Coding Agents?