What is AWS Strands - the "WordPress moment" for AI agents

Strands is AWS's free, open source toolkit for building AI agents, and since September 2026 a ready-made one: Strands Harness (Apache 2.0).

What an AI agent is

Classic software follows instructions written in advance. An agent gets a goal, a model and tools, then decides its own steps: which tool to call, in what order, when the work is done. That is why it is called the WordPress moment for agents — you no longer build the whole system yourself.

Strands and the Harness

Strands Agents is the toolkit: an open source Python SDK from May 2025, later ported to TypeScript, plus Strands Labs in February 2026. Before it, an agent meant writing the loop, tools, memory and safety rules from scratch; now a package install (pip install strands-agents) and tens of lines replace hundreds.

Strands Harness followed on 21 September 2026: assembled and tuned out of the box, installed with pip install strands-harness, with prompt caching and context management as defaults.

What Harness includes, and what it costs

It reads, writes and edits files, uses a shell and the web, manages its own context window, keeps long-term memory and sessions, and delegates subtasks; MCP servers plug in (MCP is a standard connector for external tools).

It is local-first and model-agnostic: only the model call leaves your machine, and one line swaps Amazon Bedrock for Anthropic, OpenAI, Google or local Ollama. "No features require Bedrock," says AWS's Marc Brooker. It deploys to AWS, GCP, Azure, Cloudflare or any Linux container, and its CLI exports the prototype as Python or TypeScript.

AWS puts it 45% cheaper than Claude Code and Codex at comparable accuracy across six benchmarks, or 28% including DeepSeek's cheaper harness; on Terminal-Bench 2.1 it cost $56.29 against $248.05 over 89 trials, scoring 69.7 to 61.8. Context management explains the saving: huge tool outputs are truncated, context is compacted past a threshold, overflow recovered mid-loop.

What you can build

An agent that writes a research report: define one small custom tool (a currency-rate lookup), give it the goal, and it writes report.md itself.

An agent you don't trust with your database: declare a read-only surface — queries pass, writes are blocked; the blocked action is cancelled and the agent re-plans with the access it has.

A daily briefing agent: each morning it gathers 24 hours of news, keeps a source beside every claim and returns a cited summary.

These are not toys: Swisscom and Smartsheet already run Strands in production. Memory, context and safety are packaged now, so agent building is no longer specialist-only work. That is the WordPress moment.