Agents, MCP, and Guardrails — tooling & safety for outcome engineering

llm-mcp-client 0.1a0 lands as a lightweight client for the Model Context Protocol, enabling stateless propagation of context between LLMs and tools. This makes stateless context a practical building block for agents and tools, simplifying coordination and traceability for outcome-oriented workflows (Principles 06 & 11).

Runtime: MCP goes stateless; Baseten courts lab partners reports MCP’s shift to a stateless request/response core and Baseten’s Model Labs for closed-weight model distribution. The move standardizes agent transport and lowers infra friction for integrating external models, changing how you design orchestration and deployment pipelines (Principles 07 & 11).

datasette-agent 0.4a0 adds a browser_task so agent tools can execute custom JavaScript directly in users’ browsers. That gives agents a lightweight execution surface for UI-driven tasks and local interactions, but forces you to design secure sandboxes and observable, user-consent flows (Principles 07 & 11).

Bloom Security emerges from stealth with $20M seed to monitor AI agents and extensions launches focused endpoint monitoring for AI agents, browser extensions, and plugins. Treat agent telemetry and extension surveillance as first-class security signals — you’ll need runtime detection, anomaly rules, and incident playbooks to keep agentic systems safe (Principles 14 & 15).

A Safe Path to Open Weights sets out a staged, evidence-driven framework for safely releasing open-weight models alongside strengthened ecosystem defenses and safety testing. Use staged releases, audit tooling, and community-run tests as part of your release gate to balance utility and risk when you ship models or enable third-party agents (Principles 14, 15 & 16).