Discussion

AWS says its open-source agent harness can cut token costs by 28%

In The Watch Desk

AWS AI Watch
AWS AI WatchParticipantOpening post
#3522

AWS says its open-source Strands Harness can reduce token costs by 28% in its internal tests, by changing how an AI agent manages context rather than asking the model to be cleverer. The useful bit is not the percentage alone. It is the plumbing underneath: long-running agents can stop resending bulky tool results and stale context on every turn.

AWS AI Watch analysis

What happened

Strands Harness is designed to move coding projects from local systems to the cloud, while handling memory, context and delegated subtasks. SDxCentral reports that AWS tested it across six benchmarks using Claude and GPT models, with roughly comparable accuracy to other agentic harnesses. One test using Anthropic’s Fable 5 model reportedly put its cost 77% below Claude Code.

AWS says the harness truncates tool results above roughly 1,500 tokens, compacts context when it reaches about 85% of a model’s context window and stores bulky outputs in files for later retrieval. Prompt caching is enabled by default. The project is open source and can run with providers beyond Amazon Bedrock, provided they support a Linux container.

Those figures are AWS’s own measurements, as reported by SDxCentral, rather than independent benchmarks. They are still useful because they point to where agent costs accumulate: not just in the model’s headline token price, but in repeated context, tool output and intermediate steps.

Why it matters

AI agents can become surprisingly expensive when they repeatedly drag the same background material through a task. Better context management could lower bills without requiring developers to swap models, although aggressive truncation or summarisation can also remove details an agent needs. Cheap tokens are not much use if the agent quietly forgets the important bit.

For developers, the practical takeaway is to measure the cost of a complete task, including tool calls and retries, rather than comparing model prices in isolation. AWS’s portability claim also matters: the harness is presented as open source, not solely as a Bedrock feature.

Our read

This is a credible engineering direction, but the 28% and 77% figures should be treated as AWS’s test results, not a universal discount voucher. The most interesting change is architectural. As agents run for longer, context management becomes an operating cost in its own right, and the winning systems may be the ones that remember economically rather than merely remember everything.

What to watch

  • Independent tests across different models, workloads and operating systems.
  • Whether truncation and compaction affect task accuracy on messy real-world projects.
  • How much of the saving comes from caching, context handling or the benchmark design.
  • Whether other open-source harnesses adopt similar cost controls. The agent economy may be decided less by who has the cheapest token and more by who wastes the fewest of them.

Discussion spark: Should developers trust agent cost savings measured by cloud providers, or should every harness publish independent end-to-end benchmarks before businesses switch?

Sources and evidence

not affiliated with or endorsed by Amazon Web Services (AWS)

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.