Bonsai 27B is a new quantized language model that achieves 27B-class capability—including multi-step reasoning, tool calling, and vision tasks—at 3.9 GB (1-bit variant) and 5.9 GB (ternary variant), making it the first model of this class to run on phones like the iPhone 17 Pro. The 1-bit version retains 90% of full-precision baseline performance across 15 benchmarks while delivering 10x better intelligence density than full-precision models, with critical capabilities for agentic workflows (math, coding, tool-calling) nearly untouched. This enables new deployment patterns: on-device agents with zero marginal cost per step, privacy-preserving local execution, and hybrid cloud-local architectures that route routine tasks locally while reserving expensive cloud models for frontier work.
Comments
The article breaks down agent architecture into three distinct loops: the **Inference Loop** manages chat completion API calls and maintains conversation history; the **Tool Loop** handles tool calls inferred by the LLM, including defensive checks for hallucinated tool names and tracking tool IDs; and the **Human Loop** provides approval/denial gates for tool execution, typically requiring durable execution frameworks like Temporal to handle async human decisions. These three loops—not a single monolithic agent loop—comprise the actual building blocks for autonomous agents, with the human loop being architecturally the most complex due to blocking I/O and server restart concerns.
DOOMQL is a Doom-like game built with GPT-5.6 Sol where SQLite serves as the complete game engine—handling movement, collision, enemies, combat, and rendering via SQL queries and recursive CTEs rather than traditional code. The implementation uses Python and SQLite, with a playable terminal version available on GitHub, and can be visualized in real-time through a Datasette App (built with Claude Fable) that displays game pixels and includes a minimap. This demonstrates practical use of LLMs for generating complex SQL logic and UI code, as well as novel architectural approaches to game development.
Agnost AI is a YC S26 tool that analyzes production agent conversations to detect failures that standard evals miss, identifying where users get stuck or frustrated, and automatically generating reviewed PRs to fix identified issues. It works with any LLM/framework via 2-minute OpenTelemetry-native setup, surfacing failure categories like broken workflows, repeated retries, and churn risk while also extracting feature requests from conversations. Pricing starts free (1K messages/month) with Pro at $499/month (100K messages) and custom Enterprise plans.
Willison examined GitHub's code-frequency chart for his Datasette project and observed a notable spike in coding activity at the end of the period, which he attributes to the release of newer AI models including Opus 4.8, GPT-5.5, Fable 5, and GPT-5.6 Sol. He was investigating whether coding agents and advanced language models have had a measurable impact on his own development output. The increased frequency of code changes suggests these newer models may be contributing meaningfully to his project's development velocity.