The Strands Agents team at AWS published Strands Harness on September 21 under Apache 2.0, for Python and TypeScript, deployable on a laptop or on any of five clouds. Most notable is not what the benchmark chart claims but what AWS left on it: DeepSeek Harness, the one entry that ran on fewer tokens than its own.
Key takeaways
- Strands Harness is Apache 2.0 for Python and TypeScript, instantiated by one
create_harness()call, and routes to Bedrock, Anthropic, OpenAI, Google, LiteLLM or a local Ollama model. - Three context-management defaults β 1,500-token tool-result truncation, compaction at 85% window usage, and in-loop recovery on overflow β are what AWS credits for its results.
- DeepSeek Harness consumed roughly 14% fewer tokens than the AWS loop while scoring lower on every evaluation in the set, which is why AWS's averaged saving lands at 28% rather than 45%.
Why the same release carries two savings figures
A harness is everything wrapped around the model β the loop, the tool definitions, context handling, memory and recovery when a run goes sideways. None of that is the model, and all of it decides the bill. AWS put its version against Claude Code, Codex, oh-my-pi, OpenCode and DeepSeek Harness, averaging six evaluations on EC2 through Harbor, the framework the Terminal-Bench authors built.
Restricted to the two agents most developers actually prototype in, the saving is the 45% figure The New Stack led with. Widen the field to include DeepSeek's harness and the average falls to 28%, because that one entry is cheaper than anything AWS built. Disclosing it rather than trimming the chart is a choice worth noting; it also sets up the more honest reading, which is that the cheapest harness in the field was also the least accurate one.
The comparison that isolates the variable
Benchmark tables that vary both model and harness cannot attribute anything. The Terminal-Bench 2.1 run fixes the model β Claude Fable 5, 89 trials each β and only then do the differences mean something. Against Claude Code the AWS loop was 77% cheaper and 7.9 points better. Against oh-my-pi it landed on the same 69.7 accuracy for substantially less money. DeepSeek's entry saved more and gave up 10.2 points doing it.
For a team running agents continuously, a 28% average is not a rounding difference. On a workload billing five figures a month in tokens, that is the kind of line item that justifies a migration on its own, independent of any accuracy argument.
Three defaults, and they are auditable
AWS attributes most of both the savings and the accuracy to how the harness manages context rather than to anything clever in the loop.
| Mechanism | Trigger |
|---|---|
| Tool-result truncation | Results beyond roughly 1,500 tokens |
| Compaction (summarisation) | Context usage above 85% |
| In-loop context recovery | Window overflow mid-run |
These are ordinary engineering decisions, and that is the point: the gains come from defaults any team could have set and mostly did not. A single create_harness() call returns an agent already carrying shell, file and web tools, prompt caching, memory that resumes from a session ID, and a helper agent that absorbs open-ended subtasks against a checklist. Model calls default to Amazon Bedrock and repoint to Anthropic, OpenAI, Google, LiteLLM or a local Ollama model, with MCP servers available as external tools.
Independent research points the same way
The HarnessTax study, which compared Claude Code, Codex CLI and Pi over seven models, found harness choice barely shifted success rates while the same model could reach comparable outcomes at up to five times the cost. That is an uncomfortable finding for anyone selling a harness on capability, and it happens to support AWS's pitch: if accuracy is roughly fungible, spend is the only axis left. A follow-up paper on the Strands benchmarks is promised.
It is also AWS's second open agent-infrastructure release in quick succession, following the benchmark it published without any accompanying scores. This release supplies the numbers that one withheld.
What to watch
Because the harness is a library dependency rather than an application, the agent built on a laptop is meant to be the same artifact deployed to AWS, Google Cloud, Azure, Cloudflare or Modal. A companion CLI, itself built on the harness, turns a plain-English description into exportable Python or TypeScript. The open question is whether the cost gap holds on task suites AWS did not select β the reason its promised follow-up paper matters more than the launch chart.
FAQ
Is Strands Harness free to use?
Yes. It is Apache 2.0 licensed for Python and TypeScript and distributed through GitHub and PyPI. It can run entirely offline against a local Ollama model; model API charges apply separately when a hosted provider is used.
Does Strands Harness only work on AWS?
No. Amazon Bedrock is the default model route and can be swapped in one line for Anthropic, OpenAI, Google, LiteLLM or a local model. Supported deployment targets include Google Cloud, Azure, Cloudflare and Modal.
Is it a coding agent like Claude Code?
AWS describes it as general-purpose rather than coding-specific, even though its benchmark rivals are coding agents. It arrives with shell, file and web tools and hands open-ended subtasks to a built-in helper agent.
