4,733 passing tests support a large agent surface
Our TypeScript-side run passed 4,733 tests with 0 failures, which is the best argument for taking Strands seriously. The SDK owns far more than a model call. It provides execution limits, structured output, tool hooks, MCP clients, sessions, memory, streaming, tracing, guardrails, evaluations, and several multi-agent patterns. Both Python and TypeScript implementations live in the same repository, with a shared documentation site and governance material beside them.
That breadth solves a real engineering problem. A small agent loop usually gains cancellation, retries, context management, provider accounting, approval pauses, persistence, and traces one incident at a time. Strands already has named places for those concerns. It also runs in your process without requiring a hosted control plane. The trade is commitment: once application behavior depends on its event loop and lifecycle events, changing frameworks becomes an architecture project.
The 398 MB install covers only the TypeScript path
Our sandbox installed 443 npm packages in 22 seconds and used 398 MB on disk. The root build finished in 24 seconds. TypeScript users need Node.js 22 or newer, while the Python package supports Python 3.10 through 3.14 and uses Hatch for development. The monorepo's root npm scripts point at strands-ts, so our clean result should be read as evidence for that workspace, not as a substitute for the Python matrix.
The quick start also assumes Amazon Bedrock. You need working AWS credentials and access to the documented Claude Sonnet model before the default agent can answer. The SDK supports Anthropic, OpenAI, Gemini, Ollama, and other providers, but each changes credentials, optional dependencies, response shapes, and sometimes feature behavior. Provider choice is portable at the interface level. It still deserves an integration suite using the model and region you plan to run.
What happened when we ran it
Our run at commit e7fd22b installed 443 packages in 22 seconds, built in 24 seconds, and completed tests in 88 seconds. Vitest reported 4,733 passed and 0 failed out of 4,733. The fresh unprivileged container had 3 CPUs and 8 GB of RAM. Dependency installation occupied 398 MB, while the checked-out repository itself was 26.5 MB.
The repository contained 2,547 files and about 395,625 lines of source in our scan. We found 38 CI workflow files, npm workspaces, no Dockerfile, and no directory named tests. Npm audit reported 11 known vulnerabilities, split into 3 high and 8 moderate, with none rated critical or low. Those numbers describe the measured Node environment only; they do not report Python dependency advisories or model-service behavior.
Two open bugs affect tool approval and replay
Issue 4371 was opened on September 16, 2026 against Python v1.56.0 and current main. Its reproduction says a handler calling event.interrupt() can be treated as an ordinary handler error. With on_error='proceed', the sensitive tool ran without the requested approval. The default throw policy and the separate Confirm action were not affected in that report, and paired fix pull requests were already open the same day.
Issue 4338 covers a different failure around completed tools. In its reproduction, a tool finishes, an AfterToolsEvent callback raises, and the result is not appended to conversation state. Resuming the agent runs the completed tool again. Payments, messages, orders, and database writes can all be harmed by duplicate execution. Until the fix lands in the version you deploy, give those tools idempotency keys and make recovery check external state before retrying.
Python v1.56.0 and TypeScript v1.18.0 move separately
Python v1.56.0 and TypeScript v1.18.0 were both released on September 15, 2026. Separate tags are useful because the SDKs can ship on their own cadence, yet they also make parity a versioned question. The issue queue includes explicit port requests between languages, and the latest TypeScript release notes include changes copied from Python. Check the feature page for your chosen language rather than assuming a sibling implementation behaves identically.
The repository was pushed on September 17, the day we checked it. GitHub showed 7,313 stars, 499 open issues, and 260 open pull requests. That is an active project with an unusually busy public queue, not 759 confirmed bugs. The written lifecycle policy tells production users to pin minor versions for MCP, A2A, OpenTelemetry conventions, and experimental modules because those areas may change faster than the core compatibility promise.
443 packages make sense only if the controls get used
The 443-package install is hard to justify for a thin chatbot that calls one provider and two functions. Strands earns that cost when the application needs budgets, cancellation, sessions, MCP, traces, provider switching, structured results, and operator hooks together. The Apache-2.0 license and in-process design keep deployment choices open, while 38 CI workflows and 4,733 passing tests show serious maintenance work behind the API.
Adopt it with a pinned SDK version, a real provider test, and explicit rules for tool side effects. Keep approvals on the documented safe path, make irreversible tools idempotent, and review the 11 npm advisories in the dependency paths you ship. The measured suite says the TypeScript foundation is dependable. The two open tool-control reports say your application still owns the final safety boundary.