AWS open-sources Strands harness for agent loops, claiming 28% fewer tokens

single source· 1 articles · confidence: low · first seen 2026-09-21 21:42 UTC

What this means for you

If you are hand-rolling an agent loop, this is worth a look: it is open source and runs where your code runs. Ignore the 28% figure until you reproduce it — no dataset, baseline or date is published. Nothing to migrate; it is a new component, not a replacement.

AWS's Strands Agents team has released Strands harness, an open-source scaffold for building agents — the loop that calls a model, runs its tools and carries state between steps. It runs on a local machine or deploys to a cloud provider, and supports Python. MarkTechPost reports a 28% cut in token use at comparable accuracy; no evaluation date, dataset or baseline accompanies that figure. The stated target is agents that work in Claude Code or Codex but break when developers rebuild the loop. This rests on one truncated article.

Key facts

  • ·Strands harness was released by AWS's Strands Agents team as an open-source, general-purpose agent harness. source
  • ·The reported saving is 28% in token cost at comparable accuracy, with no evaluation date, dataset or baseline given. source
  • ·It runs locally or can be deployed to a cloud provider, and ships for Python. source
  • ·MarkTechPost published the account on 21 September 2026; its article text is truncated after the Python support line. source

What the sources say

  • MarkTechPost — Positions the release as a fix for agents that break when developers rebuild the loop.

Sources

The original reporting. Follow these — they did the work.

← the wire