AWS open-sources Strands, an agent harness

AWS released Strands, an open-source agent harness. It bundles shell, file, and web tools with context management, prompt caching, long-term memory, and delegation to helper agents.

The number that matters is cost. AWS claims Strands runs 28% cheaper in tokens than competing harnesses across six benchmarks on identical models. Against Claude Code specifically, running both on Fable 5, AWS claims a 77% cost cut while Strands scores higher on Terminal Bench 2.1. A harness that cuts token spend by a quarter or more, without touching which model you call, changes the math on running agents at volume instead of hunting for a cheaper model.

It installs with pip install strands-harness or npm install @strands-agents/harness. No model swap required, since it wraps whatever you're already calling.

These are AWS's own benchmarks. The company selling the harness ran the comparison against a named competitor, and no third party has reproduced the numbers yet. Treat 28% as a starting point for testing, not a settled fact. The 77% figure against Claude Code deserves the most skepticism, since it's the number built to headline.

harness-cli v0.1.2 followed on September 23, swapping in MCP client 2.0. A fast point release a day after launch suggests AWS is planning to keep iterating rather than treating this as a one-time drop.

If token cost is what's capping how much agent work you can run, benchmark Strands against your current setup yourself before trusting AWS's numbers.