Noon roundup for September 19. Attention in this window focused on the "practical speed" of local LLMs. The open-source inference engine Inco Splash announced that it runs Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro. A hands-on tuning write-up on running a coding agent on two RTX 3090s also drew attention, sharing settings that cut average wait time per turn from 28 seconds down to 7. On the dev tools side there was smaller news too: Claude Code added AGENTS.md support, while guides on picking Codex models and on the agent-to-agent protocol "A2A" landed on Qiita.
Qwen3.8-27B at 144 tok/s — New Engine "Inco Splash" for Apple Silicon
The open-source inference engine Inco Splash for Apple Silicon is turning heads in [the LocalLLaMA community on Reddit](https://www.reddit.com/r/LocalLLaMA/comments/1wk9hze/qwen3827b_at_144_toks_on_an_m5_max_macbook_pro/). It runs Qwen3.8-27B at 144 tok/s (tokens per second) on an M5 Max MacBook Pro, and the engine itself is built for both this model and Apple Silicon. According to the developers' announcement, decode speed is up to 3x that of Ollama and 2x that of oMLX, approaching 4x in scenarios where an agent delegates work to sub-agents.
Requirements are Apple Silicon M3 or later, macOS 26.4 or later, and at least 36GB of memory. Install it from Homebrew with `brew install incoai/tap/splash` and start it with `splash serve --model incoai/Qwen3.8-27B-Splash`. The intended setup is to point coding agents such as Claude Code, OpenCode, Codex, and Hermes at it directly, and the same engine is also available inside the LM Studio app.
The numbers come from the developers' own announcement, and independent verification is still ahead. Even so, one thing is certain: the options for running a 27B-class model locally at speeds that hold up for agent workloads keep growing. Much like the expert streaming on a 64GB Mac covered in our midday article, the "practical-speed local LLMs on the Mac" direction has gained another layer of substance.
A Coding Agent on Two RTX 3090s — Tuning Wait Time from 28s to 7s
When you put a local LLM to work as a coding agent, the bottleneck often isn't the model's generation speed but how the agent is run. In [a hands-on post on Reddit](https://www.reddit.com/r/LocalLLaMA/comments/1wk8jef/tuning_qwen_38_27b_and_omp_as_a_coding_agent_on_2/), the author ran the open-source coding agent Oh My Pi (OMP) with vLLM on two RTX 3090s and shortened the average wait per turn from 28 seconds to 7 seconds.
What did the work was mostly agent-side configuration: explicitly setting the effort level for every role, since roles left unspecified fall back to the default (xhigh); capping the thinking-token budget at 7,500; raising the output limit from 8k to 32k (to keep file writes from getting cut off); offloading tool outputs over 10KB to files; limiting sub-agents to at most four and enabling appendOnlyContext. Stacked together, these changes made perceived speed roughly 4x faster. The details are also summarized in [the author's blog post](https://doug.sh/posts/tuning-a-local-coding-agent-oh-my-pi/).
It's a valuable real-world example of improvement achievable purely through operational settings, without swapping out the model itself. If an agent running in your local environment feels "slow", it seems worth reviewing where settings can be tightened before pointing fingers at the hardware.
Claude Code Now Reads AGENTS.md — Though Not in the First Session
The changelog for Claude Code version 2.1.277 (dated September 18, 2026) includes the addition of AGENTS.md support. According to [an explainer on Qiita](https://qiita.com/suwa_nobu/items/c795cf89d0fd4091c9cb), Claude Code now reads AGENTS.md when no CLAUDE.md exists in the project. AGENTS.md is an instruction file increasingly shared across multiple agent tools, which makes this meaningful for anyone wanting to avoid double bookkeeping against the Claude Code-specific CLAUDE.md.
In the author's own testing, however, it is not loaded during a project's first session. If you're migrating, don't immediately conclude the file is at fault just because the instructions don't take effect in that first session.
How to Choose a Codex Model — Splitting Astra, Sol, and Terra
Create a new task in OpenAI's coding agent Codex and you're met with three models: Astra, Sol, and Terra. [A guide posted on Qiita](https://qiita.com/inoyu-qiita/items/ab096a72824d31c154ea) sorts out when to use which, written to answer the developer's question of "I get how Claude's Fable, Opus, and Sonnet differ, but Codex is a total mystery to me."
Just yesterday this blog covered Astra as a specialized model for legal work. OpenAI's model lineup is branching into parallel lines for coding as well, and choosing per use case is now on the developer's shoulders.
Beyond MCP — Google's Agent-to-Agent Protocol "A2A"
As agents find their way into every corner of the enterprise, the next challenge is agents coordinating with each other. [An explainer on Qiita](https://qiita.com/kuno_/items/2cdde39d0fc453dec623) lays out how A2A (Agent2Agent), an open protocol that originated at Google, works.
Agents built on different vendors and frameworks can't talk to each other as-is. MCP handles connecting models (agents) to tools and data, while A2A is positioned to handle coordination between agents; the article walks through the relationship between the two and the protocol's structure in the context of Gemini Enterprise. The more agents multiply, the more standards for "how they connect" matter. For companies weighing adoption, it seems worth getting a firm grasp of the division of roles between MCP and A2A.
AI "Tilly Norwood"'s Press Tour Is Off to a Rough Start
Overseas, an AI called Tilly Norwood is making the rounds of news outlets on a press tour. According to [TechCrunch](https://techcrunch.com/2026/09/18/tilly-norwoods-press-tour-is-going-about-as-well-as-youd-expect-for-an-ai/), the reception has been poor. In one especially bizarre moment, it appeared to malfunction mid-interview and started speaking Chinese.
The article bites with its irony about an AI running the same playbook as a human publicity tour. Media relations for AI personalities are only just getting started, and the difficulty of verifying what their statements rest on looks likely to remain a nagging problem for both the journalists covering them and the readers.