AWS open-sources Strands agent harness, claiming lower token use

What happened

Amazon Web Services has released Strands, an open-source agent harness that it says consumes fewer tokens than competing agent tools while delivering close to the same benchmark results as Claude Code and Codex.

The Register reports that the headline claims come from benchmarks covering only coding agents, and that AWS ran the comparison itself. The efficiency and parity figures are therefore self-reported and have not been independently verified in the coverage.

The release is positioned as a bundling of existing components into a single open-source framework, rather than a new foundation model. The emphasis is on how the system is assembled and how economically it runs.

Why it matters

An agent harness is the orchestration layer that sits around a language model. It manages planning, tool calls, memory and the loop of attempts that lets an AI agent complete multi-step work. As agents move from demonstrations into production, the harness increasingly determines cost, speed and reliability, not just the underlying model. A claim about token efficiency is therefore a claim about unit economics: fewer tokens per task means lower running costs and faster responses for every workflow built on top.

For organisations, an open-source harness also changes the buy-or-build question. Teams can inspect how the agent loop works, adapt it to their own systems and avoid tying every workflow to a single vendor's tooling. For leaders in digital transformation, that makes the harness a design decision with consequences for governance, auditability and operating cost, not only a developer preference.

The Renascence take

The efficiency headline is the easy part to repeat. The harder question is what it means for the customer and employee experience once agents handle real service work, where tasks are messier than coding benchmarks.

Most readers will focus on the token count and miss that the benchmark was limited to coding agents and designed by the vendor. Customer service journeys involve ambiguous requests, policy exceptions and emotional stakes, and none of that is measured here. The behavioural principle at play is that cheaper interactions tempt organisations to automate more, quickly, before they have defined what a good outcome looks like. Customer-obsessed operators should pilot any harness on their own failure cases, measure resolution quality and customer effort alongside cost, and only then decide how much autonomy an agent earns.