Amazon Web Services (AWS) released an open-source agent to help developers move coding architecture from local premises to the cloud, potentially saving a pretty penny on tokens to boot.
Built on the Strands harness SDK, the Amazon offering is an agentic harness, i.e. the architecture that envelops a large language model (LLM) to turn it into a functional AI agent. Described by AWS as a general-purpose agent, Strand harness produces a ready-to-run agent when deployed, represented by frog iconography rather than the usual lobster motif when it comes to agents.
Carrying code from on-premises to cloud, the agent can handle its own context window, keep long-term memory across runs, delegate open-ended subtasks to a built-in helper agent, and load any pre-existing skills.
Interestingly, while developed by AWS, its open-source nature means it can be run not just on Amazon Bedrock but also Anthropic, OpenAI, Google, and any provider with a Linux container.
In a blog, AWS positioned Strands harness as a potential help with increasing AI token cost usage when running prompts, outputs, or intermediate reasoning steps.
Amazon’s in-house tests showed it has 28% less token cost when running the same Claude or GPT models across six benchmarks, while achieving roughly comparable benchmark accuracy to Codex and other agentic harnesses. One test, using Anthropic’s Fable 5, even suggested Strands harness costs 77% less to use than Claude Code.
Tokenomics, referring to the economics of AI infrastructure, has seen model makers like DeepSeek dramatically slash API prices in a bid to corner the frontier AI market. But as pricing drops, usage escalates, with Goldman Sachs research suggesting token usage will multiply 24-times between 2026 and 2030 to 120 quadrillion tokens per month, making for another software 'tax' enterprises need to budget for.
Under the harness hood #
Amazon put Strand's money-saving opportunity down to the tool’s automatic context management. If a tool result is larger than roughly 1,500 tokens, Strands truncates it rather than repeatedly including the entire result in subsequent model calls.
This also comes with a context compaction capability, where once active context grows beyond 85% of the model's context window, the harness triggers summarisation, replacing older detailed material with a more compact representation that preserves what the agent needs to continue working. This prevents long-running tasks from generating ever-larger prompts.
Strands harness also offloads bulky tool results into files, storing data externally where the agent can retrieve what it needs later, while enabling prompt caching by default. AWS described this as caching the reused parts of each request to save both time and cost.
Amazon stressed the tool does not "sacrifice accuracy for lower cost," claiming Strands harness "beats other agent harnesses on cost efficiency while maintaining equal or better accuracy."