Amazon Bedrock's AgentCore Harness Is GA — It Solves the One Problem That Makes Every AI Agent Team Waste Weeks

Amazon Bedrock AgentCore harness is now generally available, turning what was once a months-long infrastructure ordeal into a config file and two API calls. The announcement builds on Simon Willison's clean definition: an LLM agent runs tools in a loop to achieve a goal. While the loop itself is straightforward, the hard part has always been everything around it: wiring tools, provisioning sandboxed compute, setting up storage, secrets, networking, memory, and observability. Teams that wanted to experiment with different models or domains found themselves repeating the same plumbing. The AgentCore harness solves this by abstracting the entire production runtime — including the AgentCore primitives (Runtime, Memory, Gateway, Browser, Identity, Observability) — so developers can configure rather than build.

The harness delivers immediately: call CreateHarness to define an agent (pointing to model, tools, skills, and instructions) and InvokeHarness to run it. The agent spins up in its own isolated environment with a filesystem and shell, remembers users and conversations across sessions, browses the web, calls tools via Gateway or MCP, and streams every step in real time with automatic CloudWatch tracing. A standout capability is model flexibility: teams can plan with Claude Opus, switch to GPT-5.5 to write code, then to Gemini to summarize, all mid-session without losing context. API keys are stored securely in AgentCore Identity's token vault. New at GA: support for OpenAI GPT-5.5 and GPT-5.4 on Bedrock, plus direct access to OpenAI, Gemini, and any provider via LiteLLM. This turns agent experimentation from a heavy orchestration lift into a lightweight configuration exercise.