Amazon Bedrock AgentCore Harness goes GA

Good morning 👋 AWS made shipping a production agent a two-call process today, and while that was happening, Google retired Gemini CLI and OpenAI's reasoning model cracked 18 rare disease diagnoses that had stumped physicians.

In today's issue:

• 🔭 Amazon Bedrock AgentCore Harness is now generally available
• 🧠 Models: GPT-5.5 Instant for health, AI rare disease diagnosis, Google's Gemini CLI exit
• 🔬 Research: agent information leakage and what actually beats LoRA
• 🏛️ FERC moves on grid access for AI factories
• 🛠️ Production agent in 5 minutes: tools and how-to

🔭 THE ONE THING

🚀 Amazon Bedrock AgentCore Harness goes GA

Two API calls: CreateHarness to define the agent, InvokeHarness to run it. That's the full path to production as of today.

Each harness boots in its own isolated environment, a real filesystem and a real shell, so it can read files and run commands without you building the sandbox. AWS owns the compute, the scaling, and the isolation boundary. Agents are running in seconds, not after a sprint of container orchestration.

The piece this removes is the one most teams are still hand-rolling: a managed execution environment safe enough to give an agent real tool access. That gap has been the quiet bottleneck between "working demo" and "agent in prod." For builders on Bedrock, that layer is now somebody else's problem.

🧠 MODELS & RELEASES

• 🩺 OpenAI shipped GPT-5.5 Instant, a health-tuned model now powering ChatGPT's wellness and medical query responses, with physician-informed evaluations and stronger clinical reasoning. Health is a dedicated inference lane now, not a general-purpose afterthought.

• 🧬 OpenAI and collaborating researchers used a reasoning model to crack 18 previously unsolved pediatric rare disease cases. Real clinical files, real missed diagnoses. Not a synthetic benchmark. The distance between "AI assists medicine" and "AI finds what doctors couldn't" is closing faster than most expected.

• 🔄 Google retired Gemini CLI on June 18 and officially redirected developers to Antigravity CLI, a Go-based terminal agent co-optimized with Gemini 3.5 Flash for background workflow support. Not open-source, which has prompted community pushback, but it's the lane Google is betting on.

🔬 RESEARCH HIGHLIGHTS

• MosaicLeaks (ServiceNow) tests whether research agents leak proprietary information while completing assigned tasks. They do, more than expected. Required reading before deploying agents against internal knowledge bases.

• Beyond LoRA is Hugging Face's systematic comparison of PEFT fine-tuning methods, finding several alternatives that outperform LoRA in specific settings. If you haven't revisited fine-tuning choices recently, this is the reason to.

🏛️ POLICY & REGULATION

• FERC issued a major large-load interconnection milestone affecting how AI factory operators and semiconductor fabs can connect to the US power grid, with changes to interconnection timing and affordability rules. Material for anyone sizing compute infrastructure or looking at AI infrastructure investments.

🛠️ TOOLS & HOW-TO

Get a production agent running on AgentCore Harness in 5 minutes

1. Call CreateHarness with your agent definition: name, model, system instructions, tool list

2. Call InvokeHarness with your first input; the harness boots an isolated environment (filesystem + shell) and the agent starts immediately

3. Attach a Bedrock Guardrail config to the harness definition for real-time safety checks at any point in the flow

4. Wire external tools through the AgentCore MCP bridge; you don't manage the integration layer

Prompt: "You are a research assistant with shell and file system access. Given a topic, search available files, run any needed data queries, and produce a structured JSON report with sources, key findings, and confidence notes. Use all available tools in sequence until your report is complete."

5 trending right now

• 🎨 taste-skill gives Claude Code output genuine aesthetic judgment and stops generic slop before it ships. github.com/Leonxlnx/taste-skill
• 🔓 openclaude runs on any platform and connects to any model provider. github.com/Gitlawb/openclaude
• 📊 ChatGPT Enterprise analytics brings new spend controls and usage dashboards for org admins managing large ChatGPT deployments. openai.com
• 📡 SageMaker GenAI Insights monitors generative AI inference in real time via CloudWatch, catching latency spikes and error patterns as they happen. aws.amazon.com
• 🔍 Agentic Resource Discovery lets agents autonomously find and load models, datasets, and Spaces on Hugging Face without hardcoded lookups. huggingface.co

🔗 QUICK LINKS

• AllenAI released MolmoMotion, open weights for language-guided 3D motion forecasting from first-person camera video. Early stage, but the right direction for robotics and embodied AI builders.
• AWS published a framing piece on context intelligence for agents at scale, connecting agents to scattered enterprise data across lakes, warehouses, and unwritten institutional knowledge. More architecture than implementation, but worth the read if you're designing multi-source agent pipelines.

See you tomorrow.

Pradeep Perugu