Evaluation record · strands-agents

Strands Agents

v1.x

Amazon Web Services

Agentmodel-drivenawsopen-source

Strong

About This Agent

Open-source, model-driven AI agent SDK from AWS, used internally by Amazon Q Developer. Takes a lightweight model-first approach with MCP and A2A support, multi-agent primitives, and optional pairing with Amazon Bedrock AgentCore for hosted runtime. In June 2026 the core repo was consolidated into the strands-agents/harness-sdk monorepo (Python + TypeScript), repositioned around AWS's 'agent harness' framing; package names (strands-agents on PyPI) are unchanged.

Last Evaluated: July 9, 2026

Official Website

Trust Vector Analysis

Dimension Breakdown

🚀Performance & Reliability

task completion accuracy

Task completion testing plus review of documented internal AWS production usage

Evidence

AWS Open Source Blog — Model-driven loop proven internally by Amazon Q Developer and other AWS agent teams before open-sourcing

mediumVerified: 2026-07-09

tool use reliability

Tool invocation testing across native tools and MCP servers

Evidence

Strands Agents Documentation — Native MCP client support, decorator-based Python tools, and a maintained strands-agents-tools library

highVerified: 2026-07-09

multi step planning

Complex multi-step task testing with agent loop and graph patterns

Evidence

Strands Agent Loop Documentation — Model-driven agentic loop delegates planning to the model; workflow and graph patterns available for structured multi-step tasks

mediumVerified: 2026-07-09

memory persistence

Memory system evaluation across sessions and AgentCore Memory integration

Evidence

Strands Session Management Docs — Built-in session persistence plus integration with Amazon Bedrock AgentCore Memory for long-term memory

mediumVerified: 2026-07-09

error recovery

Error injection testing observing model-driven recovery

Evidence

Strands Agents GitHub — Tool errors surfaced back to the model for self-correction; retry and hook system available for custom recovery

mediumVerified: 2026-07-09

agent collaboration

Multi-agent coordination testing across swarm, graph, and A2A patterns

Evidence

Strands Multi-Agent Documentation — First-class multi-agent primitives: agents-as-tools, swarm, graph, and A2A protocol support

highVerified: 2026-07-09

🛡️Security

tool sandboxing

Security architecture review of SDK plus AgentCore runtime isolation

Evidence

Amazon Bedrock AgentCore — SDK itself does not sandbox tools, but AgentCore Runtime (GA Oct 2025) provides isolated sessions and a managed Code Interpreter sandbox

mediumVerified: 2026-07-09

access control

Access control assessment of IAM and identity integrations

Evidence

Strands + AgentCore Identity — Integrates with AWS IAM and AgentCore Identity for credential management and per-agent permission scoping

highVerified: 2026-07-09

prompt injection defense

Injection testing with and without Bedrock Guardrails enabled

Evidence

Strands Guardrails Documentation — Supports Amazon Bedrock Guardrails and hook-based input/output filtering; no framework-level injection defense by default

mediumVerified: 2026-07-09

data isolation

Data isolation architecture review across deployment targets

Evidence

AgentCore Runtime Session Isolation — AgentCore runs each session in dedicated microVM isolation; self-hosted deployments inherit user infrastructure isolation

mediumVerified: 2026-07-09

open source transparency

Source code and governance review

Evidence

Strands Agents GitHub — Apache 2.0, open-sourced May 2025, public roadmap, contributions from Anthropic, Meta, and other companies

highVerified: 2026-07-09

🔒Privacy & Compliance

data retention

Privacy architecture review of SDK and managed runtime options

Evidence

Self-Hosted SDK Architecture — SDK retains no data itself; retention governed by user infrastructure or AWS data policies when using AgentCore

highVerified: 2026-07-09

gdpr compliance

Compliance capabilities assessment leveraging AWS compliance posture

Evidence

AWS Compliance Programs — Deployments on AWS inherit GDPR-aligned infrastructure controls; self-hosted deployments fully controllable

mediumVerified: 2026-07-09

third party data sharing

Data flow analysis across supported model providers

Evidence

Strands Model Providers Documentation — Data flows only to the configured model provider (Bedrock, Anthropic, OpenAI, Ollama, LiteLLM); Bedrock keeps data within AWS

mediumVerified: 2026-07-09

local deployment option

Deployment options assessment including local model configurations

Evidence

Strands Ollama Provider Docs — Runs anywhere Python/TypeScript runs, including fully local execution with Ollama models

highVerified: 2026-07-09

👁️Trust & Transparency

documentation quality

Documentation completeness and accuracy review

Evidence

Strands Agents Documentation — Thorough user guide covering concepts, safety/security, deployment, and multi-agent patterns with samples repo

highVerified: 2026-07-09

execution traceability

Tracing and telemetry capabilities assessment

Evidence

Strands Observability Documentation — Native OpenTelemetry traces, metrics, and logs built into the SDK; integrates with AgentCore Observability

highVerified: 2026-07-09

decision explainability

Explainability assessment of loop transparency and hook system

Evidence

Strands Agent Loop Documentation — Model-driven loop exposes reasoning, tool selections, and intermediate steps through traces and hooks

mediumVerified: 2026-07-09

open source code

Open source assessment of license and release maturity

Evidence

Strands Agents GitHub Organization — Apache 2.0 SDK, tools, and samples; Python SDK 1.0 released 2026-05-21, TypeScript 1.0 released 2026-04-30

strands-agents/harness-sdk GitHub — Core repo renamed from sdk-python to harness-sdk (June 2026) as an Apache 2.0 monorepo shipping Python and TypeScript SDKs, CLI, and docs; old sdk-python URLs redirect

highVerified: 2026-07-09

community activity

Community engagement analysis of downloads, releases, and contributors

Evidence

AWS Open Source Blog / GitHub Metrics — 14M+ downloads by Feb 2026, rapid release cadence, and growing external contributor base

highVerified: 2026-07-09

⚙️Operational Excellence

ease of integration

Integration complexity assessment with minimal-setup testing

Evidence

Strands Quickstart — Agent in a few lines of code; model-driven approach avoids complex workflow definitions

highVerified: 2026-07-09

scalability

Scalability assessment across managed and self-managed deployment targets

Evidence

Amazon Bedrock AgentCore GA — AgentCore Runtime (GA 2025-10-13) provides serverless, session-isolated scaling; SDK also deploys to Lambda, Fargate, EKS

highVerified: 2026-07-09

cost predictability

Pricing model analysis of SDK and optional managed services

Evidence

Open Source SDK — Free Apache 2.0 SDK; costs from LLM usage and optional AWS services (AgentCore consumption-based pricing)

highVerified: 2026-07-09

monitoring capabilities

Monitoring features assessment across SDK and AWS integrations

Evidence

Strands Observability + AgentCore — Built-in OTel metrics/traces plus CloudWatch and AgentCore Observability dashboards for production monitoring

highVerified: 2026-07-09

production readiness

Production readiness assessment of API stability and documented production usage

Evidence

Strands 1.0 Releases / AWS Internal Usage — Python 1.0 (2026-05-21) and TypeScript 1.0 (2026-04-30) with semver stability; battle-tested in Amazon Q Developer

PyPI strands-agents / harness-sdk releases — Rapid post-1.0 cadence continues: Python SDK 1.46.0 (2026-07-08) and TypeScript 1.8.0 (2026-07-08)

highVerified: 2026-07-09

Strengths

+Model-driven design proven in production by Amazon Q Developer and AWS teams
+First-class MCP and A2A protocol support plus rich multi-agent primitives (swarm, graph, agents-as-tools)
+Native OpenTelemetry observability built into the SDK
+Truly model-agnostic: Bedrock, Anthropic, OpenAI, Ollama, LiteLLM and more
+Stable 1.0 APIs in both Python and TypeScript with strong adoption (14M+ downloads)
+Seamless path to managed, session-isolated runtime via Amazon Bedrock AgentCore

Limitations

!No built-in tool sandboxing in the SDK itself; isolation requires AgentCore or user infrastructure
!Deepest integrations (Guardrails, Identity, Memory, Observability) favor the AWS ecosystem
!Model-driven loop offers less deterministic control than explicit graph-first frameworks
!Younger community ecosystem than longer-established agent frameworks
!Advanced multi-agent patterns still maturing relative to the core single-agent loop

Metadata

license: Apache 2.0

supported models

0: Amazon Bedrock (Claude, Nova, Llama, etc.)

1: Anthropic API

2: OpenAI

3: Local LLMs via Ollama

4: 100+ providers via LiteLLM

programming languages

0: Python

1: TypeScript

deployment type: Self-hosted or managed via Amazon Bedrock AgentCore

tool support

0: MCP tools

1: Custom Python/TypeScript tools

2: strands-agents-tools library

3: A2A agent interoperability

github stars: 6500+ (harness-sdk monorepo, July 2026)

github repo: https://github.com/strands-agents/harness-sdk (renamed from sdk-python, June 2026)

first release: 2025 (open-sourced May 2025)

latest version: Python SDK 1.46.0 / TypeScript SDK 1.8.0 (2026-07-08)

pricing: Free (Apache 2.0) - Costs only from LLM usage and optional AWS services

python requirement: Python >=3.10

adoption: 14M+ downloads by Feb 2026; used internally by Amazon Q Developer and multiple AWS teams

Use Case Ratings

code generation

Powers Amazon Q Developer internally; strong tool use and MCP support for dev workflows

customer support

Session management, guardrails, and AgentCore identity make user-facing agents practical

data analysis

Good tool composition plus AgentCore Code Interpreter for sandboxed analysis

research assistant

Multi-agent swarm/graph patterns suit research decomposition well

financial analysis

AWS compliance posture and IAM integration help in regulated finance environments

healthcare

Viable on HIPAA-eligible AWS services but requires careful architecture review

legal compliance

Strong audit trails via OTel tracing; domain guardrails must be added