AWS team releases open-source Strands harness for lower-cost agents
AWS’s Strands Agents team has released Strands harness, an open-source system for building and deploying general-purpose AI agents. Available for Python and TypeScript under the Apache 2.0 license, it can run locally or in the cloud and is initialized with create_harness(). The system supports models through Amazon Bedrock, Anthropic, OpenAI, Google, Ollama and LiteLLM.
The harness combines an agent loop with tools, context management, memory and recovery features. Its default tools cover shell commands, file operations and web access. It can cache repeated request content, move large tool outputs into files, retain memory between runs, resume sessions and delegate subtasks to a helper agent.
In testing on six benchmarks with Harbor on Amazon EC2, the team reported 28% lower average cost than other harnesses at nearly the same accuracy. On Terminal-Bench 2.1 using Claude Fable 5, Strands harness cost $56.29 per task and achieved 69.7 accuracy, compared with Claude Code at $248.05 and 61.8. The reported efficiency gains were linked mainly to truncation of tool results above about 1,500 tokens, compaction at 85% context usage and recovery when the context window overflows.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.