AWS Strands Harness: Open-Source Agent Runtime for Python and TypeScript
AWS has released Strands Harness, an Apache-2.0 agent runtime that packages context management, persistent sessions, tools and delegation around a model-driven agent loop. The first public Python and TypeScript harness releases landed on September 21, 2026, giving developers a higher-level starting point than assembling the same capabilities directly from the Strands Agents SDK.
The public release supports Python and TypeScript. Its surrounding Strands stack supports multiple model providers, including Amazon Bedrock, Anthropic, Google, OpenAI and local models through Ollama, so the harness can be used without making the agent loop itself dependent on a single model vendor.
The practical distinction is scope. An agent SDK supplies primitives for model calls, tools and orchestration; Strands Harness supplies a configured runtime around those primitives, including context handling, session state and delegation. AWS positions that layer for agents that need to continue across longer tasks and deployment environments.
What shipped
The September 21 GitHub releases establish the first public harness packages:
| Package | First public release | Language |
|---|---|---|
harness-python |
0.1.0 | Python |
harness-typescript |
0.1.0 | TypeScript |
The repository is licensed under Apache License 2.0. Strands also exposes file, shell and web tooling, MCP integration, configurable model providers and a CLI-oriented workflow for prototyping agents.
For local-model users, Ollama support is the important deployment option. Model quality and tool-use reliability still depend on the selected model; the harness supplies the surrounding execution machinery rather than a bundled foundation model.
Context management is a core feature
Long-running agents can spend substantial tokens repeatedly carrying tool output and prior conversation state. Strands Harness includes defaults intended to control that growth. AWS describes mechanisms for truncating large tool results, compacting context as the available window fills and recovering from context overflow inside the agent loop.
That makes the harness relevant beyond convenience. Context policy can affect token consumption, latency and whether an agent can keep working through a long task without exhausting its model window.
AWS benchmark claims need workload context
AWS benchmarked Strands Harness across six agent evaluations: ALFWorld, ContextBench, GAIA, WebShop, tau3-bench and Terminal-Bench 2.1. Independent reporting on AWS's published results says the company measured broadly comparable aggregate accuracy to Claude Code and Codex while reporting lower average task cost.
The benchmark remains a vendor-run comparison. The competing systems are also not identical products: AWS describes Strands Harness as a general-purpose agent runtime, while several systems in the comparison are primarily coding agents. Cost results additionally depend on model selection, prompts, provider pricing, caching and benchmark configuration.
One useful result is narrower and easier to interpret. The published Terminal-Bench 2.1 comparison reports Strands Harness at 69.7 with $56.29 total cost across 89 trials, versus Claude Code at 61.8 and $248.05 in the same AWS test. DeepSeek Harness cost less at $40.30 and scored 59.5. These are AWS-reported measurements, not an independent benchmark reproduction.
Where Strands Harness fits
Strands Harness is most relevant when a team wants a reusable agent runtime while retaining control over the model provider and deployment target. The open-source layer can run locally or in infrastructure selected by the developer, while Amazon Bedrock AgentCore remains an optional AWS-managed deployment path.
That separation also makes the project useful for local-AI experimentation. A developer can keep the orchestration layer and switch from a hosted model to an Ollama-served model, then measure whether the local model maintains sufficient tool-use and reasoning quality for the workload.
For production evaluation, the main tests are operational: task completion rate, token and inference cost, context growth, tool-call reliability, permission boundaries, session recovery and behavior when the selected model reaches its context limit. Those measurements are more transferable than a single aggregate benchmark score.
Bottom line
Strands Harness turns AWS's existing Strands agent primitives into a more complete, open-source runtime for persistent agent work. The Apache-2.0 code, Python and TypeScript packages, swappable model providers and local-model path make it a credible option for teams that want control over both orchestration and deployment. AWS's initial benchmark results are promising enough to justify testing, with workload-specific measurements still required before choosing it over an established coding or general-purpose agent harness.