Before this week, most teams lived with two agent setups. Claude Code from Anthropic or Codex from OpenAI already felt complete on a laptop. The tools, the files, and the loop were there. The same team still had to assemble a body when the job was research, tickets, or anything that had to run off the laptop. Someone wired tools, decided how much of a search dump went back to the model, kept a session, and tried to stop the context window from filling with junk. A large share of the token invoice came from that wrap.
On September 21 the Strands Agents team at Amazon Web Services released Strands harness, a fully assembled agent you run locally or deploy in a Linux container. Swami Sivasubramanian, who leads agentic AI at AWS, posted the launch and credited Marc Brooker and the team. The license is Apache 2.0. One line of Python or TypeScript starts it. You point it at Amazon Bedrock, Anthropic, OpenAI, Google, or a model running on your own machine through Ollama. Bedrock is the default path. You change the model in that same line.
The team built it as a general-purpose agent. Out of the box it can use a shell, read and write files, search the web, keep a session identifier so work resumes, hand a side task to a helper agent, keep a checklist, and load extra skills if you add them. A command-line tool lets you describe the agent in plain English and export the code. You can put the same package on Modal, Cloudflare Containers, Azure Container Apps, Google Cloud Run, Amazon ECS, or Amazon Bedrock AgentCore.
AWS Scores the Wrap on Token Cost
AWS ran the same Claude or GPT models inside several harnesses across six benchmarks and said Strands used 28 percent fewer tokens while matching or beating accuracy against Claude Code, Codex, and other popular harnesses. DeepSeek Harness used even fewer tokens and posted lower accuracy. On Terminal Bench 2.1 with Fable 5, the team reported Strands at $56.29 and a 69.7 score across 89 trials, against Claude Code at $248.05 and 61.8. Testing ran on EC2 with Harbor. A follow-up paper is promised.
Those defaults are the mechanism. The harness caches the reused prompt so you do not pay for it on every hop. Tool results over about 1,500 tokens get cut and the rest can sit in a file. When the context window passes 85 percent, older turns get summarized. If the window overflows, the loop tries to recover. Two teams can use the same model and spend very different amounts because one of them keeps sending the whole log and the whole web page back.
The Register noted that AWS scored its own work and raced mostly coding harnesses. MarkTechPost and The New Stack published the same tables. Treat 28 percent as AWS’s published result until someone else reruns the tests.
What Changes for a Buyer This Quarter
Finance should put the token line next to the seat line. If agents are already in production, a large share of spend is repeated context. A cheaper wrap with similar answers is an operating-cost decision. Pick one real job, a research memo, a ticket writeback, or a file edit, and run it in the current stack and in Strands with the model you already approved. Decide on that receipt.
Claude Code and Codex already ship a finished body for coding on a laptop. Teams still have to choose a body for every other job. Custom agents stalled when memory and truncation were rebuilt badly. Strands is AWS publishing opinionated defaults for that layer. Override them if your job requires it. The defaults now exist in public code.
Security should treat an open harness as software that can touch a disk and a network. Shell and file tools ship unless you take them away. Ask what gets written when a fat tool result is parked in a file, who can resume a session identifier, and whether the approved model can stay off the public internet. AWS’s incentive is visible. If this body becomes common, more of those agents still call Bedrock or land on AWS compute when they leave the laptop. Useful defaults and an AWS on-ramp sit in the same package. Policy can keep them separate.
Ask finance which agent jobs already burn tokens, and how much of last month’s bill was repeated context rather than new work.
Ask engineering to rerun one approved task in Strands against the current harness, using the same model, and to bring back cost and answer quality.
Ask security which tools the agent is allowed to use, where truncated tool output is stored, and whether the model path can stay on the approved list including a local model.
Strands Agents. "Introducing Strands harness: frontier performance with 28% lower token cost." 21 Sep. 2026, https://strandsagents.com/blog/introducing-strands-harness/.
Sivasubramanian, Swami. Post announcing Strands harness. X, 22 Sep. 2026, https://x.com/SwamiSivasubram/status/2102217331006988718.
MarkTechPost. "AWS Strands Agents Team Releases Strands Harness." 21 Sep. 2026, https://www.marktechpost.com/2026/09/21/aws-strands-agents-team-releases-strands-harness/.
Vigliarolo, Brandon. "AWS bolts together open source agent harness, says it sips fewer tokens than rivals." The Register, 21 Sep. 2026, https://www.theregister.com/ai-and-ml/2026/09/21/aws-bolts-together-open-source-agent-harness-says-it-sips-fewer-tokens-than-rivals/5297915.
The New Stack. "AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex." 22 Sep. 2026, https://thenewstack.io/aws-strands-harness-agent/.
