AWS open-sources Strands Harness for LLM-agnostic AI agents

Strands Harness sits on top of the Strands Harness SDK that AWS published in August 2026, and it treats the "harness" — the surrounding machinery that gives an agent memory, file and network access, guardrails and permissions — as the part worth standardizing. That framing matters because AWS says the harness largely determines whether an agent pursues its goal efficiently, wastes tokens, or lets guardrails intervene.

The published comparison is the sharpest part of the release. Running on the same Fable 5 base model, AWS reports its own agent scoring 69.7 at $56.29, describing it as lower-cost and higher-capability than Claude Code, which it places at the bottom. Because the agent can be re-pointed at a different LLM after it is built and deployed to any Linux container environment, the competitive question is less about which model wins and more about whether the harness layer becomes a commodity teams can swap freely.

For companies already standardized on a single vendor's agent tooling, the appeal is optionality: few lines of code produce an agent with shell execution, file reading and writing, and web search out of the box, with context compression at 85% of the window and session memory. Whether that pulls teams away from LLM-locked tools is likely to hinge on independent testing of the benchmark claims and on how much customization users actually need.