AWS Strands Harness Cuts Agent Token Costs 28% on Launch
AWS released Strands Harness on September 21, 2026, an open-source general-purpose agent framework that reduces token costs by 28% compared to rival harnesses running the same model, and by as much as 77% compared to Claude Code in the same benchmark. The launch is the first major expansion of the Strands Agents project beyond a bare SDK into a complete, deployable agent ready to run out of the box.
For enterprise teams building AI agents at scale, the timing matters. Token spend on agentic workloads has become the primary variable cost in AI operations, and most enterprise AI pricing discussions in 2026 have centered on model selection rather than harness efficiency. Strands Harness argues that the framework layer matters as much as the model itself.
What Launched
Strands Harness is a fully assembled agent that wraps the Strands Agents SDK into a ready-to-run configuration. It ships under Apache 2.0, installs with one command (pip install strands-harness or npm install @strands-agents/harness), and starts with a single function call: create_harness().
Out of the box it runs on the latest models from Amazon Bedrock, Anthropic, OpenAI, and Google. Teams that want to keep inference local can point it at an Ollama model instead. It deploys to Modal, Cloudflare Containers, Azure Container Apps, Google Cloud Run, Amazon ECS, and Amazon Bedrock AgentCore without configuration changes.
Unlike Claude Code, Codex, or Kiro, Strands Harness is explicitly not a coding agent. It ships with shell, file, and web tools as its default primitives, which maps to the general range of tasks a frontier model already knows how to handle. The pitch is a baseline for any agentic workflow, not a purpose-built vertical tool.
The Benchmark Numbers
AWS ran distributed benchmarks on Amazon EC2 using Harbor, an evaluation framework from the Terminal-Bench project. The tests covered six tasks: ALFWorld, ContextBench, GAIA, WebShop, tau2-bench, and Terminal-Bench 2.1. Cost was measured as average dollars per task.
| Harness | Cost per Task | Accuracy |
|---|---|---|
| Strands Harness | $56.29 | 69.7% |
| DeepSeek Harness | $40.30 | 59.5% |
| OpenCode | $73.42 | 66.3% |
| Oh-my-pi | $86.83 | 69.7% |
| Claude Code | $248.05 | 61.8% |
Strands Harness and Oh-my-pi tied on accuracy at 69.7%, but Strands Harness cost 35% less. Against Claude Code, the cost advantage reached 77% while accuracy was 7.9 percentage points higher. DeepSeek Harness was the cheapest option at $40.30 per task, but accuracy dropped to 59.5%, which AWS noted as a consistent pattern across the evaluation.
The benchmark was conducted with Anthropic’s Fable 5 model, so the efficiency gains come from the harness layer rather than model selection.
Why the Harness Layer Matters
The efficiency gap narrows to three default behaviors built into Strands Harness. Understanding them helps enterprise teams replicate the logic in their own agent systems or evaluate whether the defaults are appropriate for their workloads.
Token truncation for tool results. Any tool output exceeding roughly 1,500 tokens gets truncated before entering the context window. Verbose outputs from file reads, shell commands, or web searches are the most common source of bloated context in production agents. Truncating them by default removes a major source of wasted spend without requiring agents to handle truncation logic themselves.
Compaction at 85% context usage. When the context window reaches 85% capacity, Strands Harness automatically summarizes and compresses earlier conversation history. This prevents the agent from accumulating irrelevant context from earlier turns, which is a common failure mode on long-running tasks where early exchanges become noise.
In-loop overflow recovery. If the context window does overflow despite proactive management, the harness recovers within the same agent loop by compacting history and continuing. Most frameworks either crash or require a manual restart; the in-loop recovery means production agents can handle edge cases without operator intervention.
SiliconANGLE confirmed the launch and noted that the efficiency gains reflect AWS’s position as an operator of production agent workloads at scale, where context management decisions have a measurable impact on monthly cloud bills.
Strategic Context for Enterprise Teams
The Strands Agents project represents AWS’s entry into the open-source agent framework market, placing it alongside AWS Kiro Crew on the coding side and now Strands Harness on the general-purpose side. The Apache 2.0 license means enterprise teams can embed, modify, and ship it without restriction, which matters for teams building proprietary agentic products on top of the framework.
The benchmark results challenge a common assumption in enterprise AI procurement: that model selection drives costs more than framework design. The same Fable 5 model costs 77% more per task inside Claude Code than inside Strands Harness. That gap is entirely attributable to how each framework manages context, not the model’s pricing per token.
For teams running agentic workloads at volume, the math compounds quickly. A team running 10,000 agent tasks per month at Claude Code’s observed rate of $248 per task would spend $2.48 million. At Strands Harness rates of $56.29, that drops to $562,900. The model costs the same per token in both cases.
The open-source agent harness market has been consolidating around a few frameworks after a period of proliferation. AWS entering with Strands Harness and backing the benchmark with production data from its own EC2 infrastructure is a meaningful signal about where the default choices will settle for teams that do not have a strong framework preference already.
What to Watch
AWS researchers have a paper forthcoming on the benchmarking methodology, which will give teams independent criteria to evaluate the claims. The benchmark was conducted by the same team that built the harness, so external replication will matter before enterprise procurement decisions treat the numbers as definitive.
Strands Harness does not include specialized tooling for domain-specific tasks such as code editing, database operations, or enterprise data connectors. Teams with specialized workloads will need to add custom tools, which is supported by the framework’s design but requires additional implementation work.
The current benchmark covers accuracy and cost, not latency or reliability under failure conditions. For enterprise workloads where response time or error handling is a primary constraint, additional evaluation against those dimensions is warranted before switching frameworks.
The Strands Harness source code is available on GitHub under the Strands Agents organization. Installation and full documentation are at strandsagents.com.