← Back to research
·10 min read·opensource

AWS Strands Agents

AWS Strands Agents is Amazon's open-source, model-driven agent SDK — the LLM plans its own loop — pulling 16M+ monthly PyPI downloads and powering Amazon Q Developer and AWS Glue internally, with the core repo recently renamed from sdk-python to harness-sdk as it consolidates into a multi-language agent harness.

Key takeaways

  • The strands-agents Python package pulls 16.7M downloads per month as of June 2026 against a comparatively modest 6,100+ GitHub stars — adoption runs through AWS's own products (Amazon Q Developer, AWS Glue, VPC Reachability Analyzer) and customer base more than through community mindshare
  • The core repo's rename from sdk-python to harness-sdk signals a repositioning: one monorepo now ships Python (v1.42.0), TypeScript (v1.4.0), and an experimental Python-WASM build, aligned with AWS's "agent harness" framing and the Bedrock AgentCore harness, which assembles Strands agents from declarative config
  • Architecture is maximally model-driven — the LLM dynamically directs its own steps and tool use rather than following hardcoded workflows — with native MCP, A2A, multi-agent graph/swarm primitives, OpenTelemetry tracing, and 12+ model providers despite the Bedrock-first defaults

FAQ

What is AWS Strands Agents?

Strands Agents is AWS's open-source, Apache 2.0 agent SDK (Python and TypeScript) that takes a model-driven approach — the LLM plans its own reasoning loop and tool use — rather than requiring developers to define explicit workflows.

How much does Strands Agents cost?

The framework is free under Apache 2.0 with no commercial tier. Costs come from the infrastructure you run it on (Lambda, Fargate, EC2, Bedrock AgentCore) and model inference, billed at standard AWS or provider rates.

What models does Strands Agents support?

Amazon Bedrock is the default (an Anthropic Claude Sonnet model in us-west-2), with first-class providers for Anthropic, OpenAI, Gemini, Cohere, Mistral AI, Ollama, llama.cpp, Llama API, SageMaker, Writer, and anything reachable via LiteLLM, plus custom providers.

How is Strands Agents different from Google ADK?

ADK pairs deterministic workflow agents with LLM-routed delegation and targets Google Cloud's managed runtime; Strands bets harder on pure model-driven planning — the LLM directs its own loop — and routes naturally to AWS deploy targets like Lambda, Fargate, and Bedrock AgentCore.

Executive Summary

AWS Strands Agents is Amazon's open-source agent SDK, launched May 16, 2025 by Clare Liguori (Senior Principal Engineer, AWS Agentic AI) — the same framework AWS uses in production for Amazon Q Developer, AWS Glue, and VPC Reachability Analyzer.[1] Its design bet is maximal model-direction: rather than hardcoded workflows, "the agent uses the model to dynamically direct its own steps and to use tools" — the LLM plans its own loop.[1] The core repo, recently renamed from sdk-python to harness-sdk, holds 6,100+ GitHub stars and 870+ forks under Apache 2.0 as of June 2026, with a separate TypeScript SDK at 700 stars.[2][3]

The headline number is downloads, not stars: the strands-agents Python package pulls 16.7M downloads per month as of June 2026 — within range of far more famous frameworks — reflecting distribution through AWS's own products and enterprise customer base rather than community buzz.[4] The rename to harness-sdk signals where AWS is taking it: one monorepo shipping Python, TypeScript, and an experimental Python-WASM build, positioned as the "agent harness" layer underneath Bedrock AgentCore's declarative harness offering.[5][6]

AttributeValue
CompanyAmazon Web Services
LaunchedMay 16, 2025 (open-sourced from internal use)[1]
FundingN/A — AWS product
GitHub Stars6,100+ core repo + 700 TypeScript SDK (June 2026)[2][3]
LicenseApache 2.0[2]

Product Overview

Strands pitches "a model-driven approach to building AI agents in just a few lines of code": define a system prompt and tools, and the model handles planning, reasoning, and tool selection autonomously — a contrast with frameworks that require "complex prompt instructions" and explicit orchestration graphs.[2][1] Agents can be conversational, event-triggered, scheduled, or continuously running.[1]

Key Capabilities

CapabilityDescription
Model-driven loopThe LLM dynamically directs its own steps and tool use; lightweight, customizable agent loop[1][2]
Multi-agent systemsGraph, swarm, handoff, and workflow primitives; multi-modal inputs in graphs since v1.0[2][5]
MCP nativeFirst-class MCP server integration for tools[2]
A2A protocolAgent-to-agent communication, shipped with the 1.0 line[5]
StreamingResponse streaming plus experimental bidirectional audio streaming[2]
Hot-reloading toolsTools loaded and reloaded from directories during development[2]

Product Surfaces / Editions

SurfaceDescriptionAvailability
Python SDKFlagship framework, strands-agents on PyPIGA — v1.0.0 July 15, 2025; v1.42.0 June 2026[5]
TypeScript SDKParallel implementation (started Sept 2025)v1.4.0, June 2026[3][5]
Python-WASMWebAssembly build of the harnessExperimental — v0.0.1, June 2026[5]
AgentCore harnessBedrock AgentCore assembles Strands agents from declarative configPreview[6]

Technical Architecture

Strands is a library, not a platform: it runs anywhere Python 3.10+ runs, from local scripts to Lambda, Fargate, and EC2, with documented production deployment architectures for each.[2][1]

pip install strands-agents strands-agents-tools

The June 2026 monorepo consolidation is architecturally telling: harness-sdk now cuts versioned releases per runtime — python/v1.42.0, typescript/v1.4.0, python-wasm/v0.0.1 — from a single repo, and Amazon Bedrock AgentCore's harness (Preview) sits on top, turning declarative configuration into an assembled Strands agent.[5][6]

Key Technical Details

AspectDetail
DeploymentLambda, Fargate, EC2, local; Bedrock AgentCore harness (Preview)[1][6]
Model(s)Amazon Bedrock (default: an Anthropic Claude Sonnet model, us-west-2), Anthropic, OpenAI, Gemini, Cohere, Mistral AI, Ollama, llama.cpp, Llama API, SageMaker, Writer, LiteLLM, custom[2]
IntegrationsMCP tools, A2A protocol, strands-agents-tools package[2][5]
ObservabilityOpenTelemetry-based tracing and metrics[2]
Open SourceYes (Apache 2.0)[2]

Strengths

  • Production provenance — Open-sourced from a framework already running Amazon Q Developer, AWS Glue, and VPC Reachability Analyzer, not built greenfield for marketing[1]
  • Massive quiet distribution — 16.7M monthly PyPI downloads as of June 2026, an adoption signal stars dramatically understate[4]
  • Purest model-driven design — The LLM plans its own loop; as frontier models improve at planning, Strands captures the gains without framework rewrites[1]
  • Genuinely multi-provider — 12+ providers including OpenAI, Gemini, Mistral, Ollama, and llama.cpp, despite the Bedrock default[2]
  • Open-protocol posture — Native MCP for tools and A2A for inter-agent communication rather than proprietary equivalents[2][5]
  • Fast, steady cadence — v1.0.0 in July 2025 to v1.42.0 by June 2026, roughly a release per week on the stable line[5]
  • Expanding runtime surface — TypeScript at parity-track v1.4.0 and an experimental WASM build broaden where Strands agents can run[5][3]

Cautions

  • Thin independent community — The HN launch thread drew 10 points and zero comments, and no Strands submission has cracked the front page since; discussion lives almost entirely on AWS-owned channels[7]
  • Stars lag the category — 6,100+ stars versus 20,000+ for Google ADK and 17,600+ for Pydantic AI; the download number likely includes heavy AWS-internal and CI-driven installs, so treat it as an upper bound on human adoption[2][4]
  • Branding churn — sdk-python renamed to harness-sdk, "Strands" versus "AgentCore harness" overlap, and per-runtime release tags make the project harder to track than its competitors[5][6]
  • AWS gravity — Defaults assume Bedrock in us-west-2, and deployment guidance centers on Lambda/Fargate/EC2; portable in principle, AWS-shaped in practice[2][1]
  • Issue backlog — 648 open issues against the core repo as of June 2026 for a framework of this size[2]
  • Model-driven tradeoff — Letting the LLM direct its own loop means less determinism than ADK-style workflow agents; teams needing auditable, fixed pipelines must build that discipline themselves[1]

Pricing & Licensing

TierPriceIncludes
Strands Agents (all SDKs)FreeFull framework, tools package, MCP/A2A support (Apache 2.0)[2]
AWS deploymentUsage-based AWS pricingLambda, Fargate, EC2, Bedrock model inference[1]
Bedrock AgentCore harnessUsage-based AWS pricing (Preview)Managed assembly and runtime for Strands agents[6]

Licensing model: Open source (Apache 2.0); AWS monetizes through Bedrock inference and compute consumption, not the framework.[2]

Hidden costs: Model inference dominates — the default Bedrock Claude configuration bills per token, and third-party providers bill separately. Observability backends and AgentCore preview pricing are additional and workload-dependent.[2][6]


Competitive Positioning

Direct Competitors

CompetitorDifferentiation
Google ADKADK mixes deterministic workflow agents with LLM delegation and has 3x the stars; Strands is more purely model-driven and AWS-native rather than GCP-native
Microsoft Agent FrameworkThe Azure mirror image; choose by cloud allegiance — Strands wins inside AWS shops
Pydantic AIPydantic AI leads on type-safe validated outputs and durable execution; Strands leads on AWS deploy paths and internal-production pedigree
LangChainLangChain has the largest ecosystem and community; Strands is a smaller, more opinionated harness with cleaner model-driven defaults

When to Choose Strands Agents Over Alternatives

  • Choose Strands when: You run on AWS, want the same framework that powers Amazon Q Developer, and prefer letting the model plan its own loop over wiring explicit graphs
  • Choose Google ADK when: You live on Google Cloud or need deterministic workflow agents alongside LLM delegation
  • Choose Pydantic AI when: Type-safe structured outputs and durable execution (Temporal/DBOS) are the priority
  • Choose LangChain when: You need the deepest integration catalog and community knowledge base

Ideal Customer Profile

Best fit:

  • AWS-native engineering teams deploying agents to Lambda, Fargate, or Bedrock AgentCore
  • Organizations that want a framework with proven internal production use at Amazon scale
  • Teams betting on model-driven planning over hand-built orchestration graphs
  • Builders standardizing on open protocols (MCP, A2A) for tools and agent interop
  • Python-first shops, with TypeScript teams now viable on the v1.x TS SDK

Poor fit:

  • GCP- or Azure-committed organizations
  • Teams that need deterministic, auditable, fixed-step pipelines as the default
  • Projects that depend on a large community of tutorials, plugins, and Stack Overflow answers
  • Developers wary of AWS naming churn and preview-stage managed offerings

Viability Assessment

FactorAssessment
Financial HealthBacked by AWS — no funding risk; monetized via Bedrock and compute consumption
Market PositionQuiet heavyweight — 16.7M monthly downloads but 6,100+ stars; distribution over mindshare[4][2]
Innovation PaceRapid — 42 minor releases on the 1.x line in 11 months, plus TS and WASM runtimes landing[5]
Community/EcosystemWeak independent community — AWS-channel-centric, near-zero HN/Reddit footprint[7]
Long-term OutlookPositive — it anchors AWS's agent strategy and underpins the AgentCore harness[6]

Strands' viability question is the inverse of most small frameworks: the code and the backing are not in doubt — Amazon Q Developer and AWS Glue run on it — but the community is. The harness-sdk rename and AgentCore harness integration suggest AWS sees Strands less as a standalone open-source brand and more as the substrate layer of its managed agent stack, which is good for longevity and ambiguous for open-source governance.


Bottom Line

Strands Agents is the AWS answer in the agent-framework wars: a clean, model-driven harness with real production pedigree, open protocols, broad provider support, and the smoothest path to AWS deployment — carried by enterprise distribution rather than developer buzz. The 16.7M-downloads-versus-6,100-stars gap is the whole story in one stat.[4][2]

Recommended for: AWS-committed teams shipping production agents who want Amazon's own internally proven framework, model-driven autonomy, and native MCP/A2A support.[1][2]

Not recommended for: Multi-cloud purists, teams that need deterministic workflow primitives as defaults, or developers who rely on a large independent community for support.

Outlook: The rename from sdk-python to harness-sdk, the per-runtime monorepo releases, and the AgentCore harness preview all point one direction — Strands becoming the open substrate of AWS's managed agent platform.[5][6] Watch whether the TypeScript and WASM runtimes broaden adoption beyond the AWS faithful, and whether independent community traction ever catches up to the download curve.[3][8]


Research by Ry Walker Research • methodology