Google's 2026-07-28 MCP release candidate removes transport-level session management, killing the initialize handshake and session-ID pinning. That unblocks horizontal scaling, serverless deploys, and round-robin load balancing for agent servers, in what Google calls the biggest MCP change since launch.
Simon Willison built a lightweight web chat UI for exercising any OpenAI-Responses-compatible endpoint, aimed at browser-based testing of local LLMs (via LM Studio's --cors and OpenRouter) without hitting CORS walls. It persists conversations client-side and even streams in generated SVG images as they render, all built with GPT-5.6-Sol xhigh.
jit moves secrets from plaintext files (.env, ~/.aws/credentials.npmrc, shell exports) into a local Touch-ID-gated vault, rewriting files to decoys and injecting real values only into the process that asks. It's a smart answer to AI agents and package installers running with your full permissions. macOS-only (Apple Silicon) and still in development.
Researchers used an AI agent to GPU-port CReSS, a 250,000-line legacy Fortran weather code, validating each kernel against dumps from a real typhoon run. It produced 162 verified GPU kernels at 5.1x speedup while catching numerical discrepancies in five. The takeaway: AI-assisted porting demands validation-centric workflow design, not just code generation.
Michael Stapelberg spent years losing commands from ~/.zsh_history before tracking down the culprit: a truncation bug fixed in Zsh 5.9.2. His winning move, patching Zsh to crash loudly and analyzing the core dump, is a masterclass in debugging rare data-loss bugs.
Dmitry Grinberg argues RISC-V should have known better: in embedded, its one natural market, the ISA is clearly worse than incumbents on code density and interrupt latency. Compressed-offset limits and late extensions like Zba mean it trades away precisely what low-cost microcontroller cores need.
LittleLearner trains 0.6B, 5B models from scratch on an 88B-token corpus filtered to U.S. elementary-school content, with matched unfiltered controls. Scaling, RL, and in-context learning amplify in-scope skills, but none push capability past the K, 5 curriculum, suggesting pretraining data sets the true capability ceiling.
Anthropic's first survey of emerging multi-agent systems shows individual agent quirks, like confabulation and reward hacking, compounding into unexpected systemic failures at scale. With agent-agent interaction volume set to outpace human interaction, the research argues we don't yet know how to make multiagent coordination go well.
Google explains why standard request-based load balancing collapses for real-time agents: long-lived bidirectional streams carry audio buffers, transcripts, tool calls, and context, so QPS and CPU miss the real load. Balance both signals by adding active-session counts to route work sensibly.
You’re caught up.
That was all 9 stories in the 16 August issue.