GitHubGitHub RepoNeeds Review

strands-agents/harness-sdk

Strands packages lifecycle limits, memory, MCP, tracing, guardrails, and evals into an in-process agent harness, with lower-level Python and TypeScript SDKs when defaults stop fitting.

GitHub · Trending today

Open Source Open Markdown Open JSON

Source Summary

Strands Agents is an **Apache 2.0** in-process SDK for Python and TypeScript with no hosted control plane. Its assembled harness includes turn and token limits, cancellation, memory, sessions, MCP, streaming, guardrails, tracing, evals, and multiple model providers.

Practical Implication

Start with **create_harness()** or **createHarness()** when you need production controls without writing the loop. Drop to the SDK when custom tools, hooks, providers, memory, or execution behavior justify owning more of the harness.

Agent-Ready Context

Strands Agents is an **Apache 2.0** in-process SDK for Python and TypeScript with no hosted control plane. Its assembled harness includes turn and token limits, cancellation, memory, sessions, MCP, streaming, guardrails, tracing, evals, and multiple model providers.

Start with **create_harness()** or **createHarness()** when you need production controls without writing the loop. Drop to the SDK when custom tools, hooks, providers, memory, or execution behavior justify owning more of the harness.

The repository describes broad portability and benchmarked defaults but supplies no comparative results in the fetched material. Provider behavior can still differ, so portability claims need testing against your tools, budgets, and failure modes.

Connected Context · Feed7 Judgment

This packages many previously separate production-harness recommendations into an open, in-process SDK with a high-level starting point and lower-level escape hatches. It supports standardizing loop controls, memory, tracing, guardrails, and provider access without adopting a hosted control plane, while leaving application-specific authority, recovery semantics, and portability testing to the builder.

Harness Engineering: Building the Production Cage for Powerful Domain Agents — Mike Chambers, AWSThe SDK is a concrete implementation of the proposed standardized harness boundary, assembling tools, memory, loop control, tracing, evaluation, and runtime safeguards into reusable components.AI Agents Are Just Distributed Systems Now — Salman Munaf, TikTokTurn limits, cancellation, sessions, and tracing cover part of the distributed-systems reliability surface, but idempotency, outcome reconciliation, compensation, and safe recovery still depend on application tool design.Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMindThis offers the opposite abstraction choice from reducing an agent to instructions and setup files: use the assembled SDK when explicit hooks and execution controls justify owning orchestration, while still starting from its high-level factory.Give the Agent a Budget, Not a Token — Sachin Malhotra, AnthropicToken and turn limits bound computation but do not provide the candidate’s quantitative limits on write authority, action rate, or irreversible effects, so those controls remain external implementation requirements.

Context Map

agentcoding#harness-engineering#agent-sdks#tool-use

Uncertainty

The repository describes broad portability and benchmarked defaults but supplies no comparative results in the fetched material. Provider behavior can still differ, so portability claims need testing against your tools, budgets, and failure modes.