Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
LFM2.5-24B-A2B: Built for Fast On-Device Agents
Liquid AI releases LFM2.5-24B-A2B, an open-weight 24B-parameter reasoning model with 2B active parameters aimed at fast, private on-device agents. The post covers changes since LFM2-24B-A2B, agentic and tool-use benchmarks against Qwen3.6-35B-A3B, inference speed on Apple and NVIDIA hardware, and how to get started.
6 min · 1,330 words
I Deleted 168,000 Lines of Comments. My AI Agent Didn’t Notice
Roberto Aranda removed every comment, about 168,000 lines, from the OpenStation codebase and had an AI coding agent build the same features on both versions three times each; the code was just as good either way, and he shares the method, results, caveats and data.
4 min · 871 words
How Cogent Runs Cloud Agents Securely
Cogent's engineering team describes how it runs cloud security agents assuming the sandbox may be compromised: trusted systems outside the sandbox control credentials, network and data access while still letting agents verify their work.
7 min · 1,553 words
arXiv 2610.18861: How can we build an agent that thinks beyond human language, yet still acts through it in the external environment? Modern LLM agents still reason and interact through text, while latent reasoning offers an efficient internal medium but lacks native tool access and reasoning transparency. We bridge these two directions with Hybrid Reasoning Agents (HybridAgents), which dynamically navigate between latent reasoning and text-based interaction throughout a unified generation process. Architecturally, a lightweight HybridLink enables seamless transitions betwe
1 min · 235 words
Lost in Implicit Routing: How LLM Agents Fail to Load Relevant Skills at Scale
arXiv paper showing skill routing collapse: as irrelevant skills are added, LLM agents fail to load the right reusable skills on SkillsBench even when golden skills stay fixed.
2 min · 399 words
Zero to Calc in Six Days: Inside XBOW’s Chrome Full Chain
XBOW walks through the Chrome full-chain exploit its AI agents built in six days, chaining a V8 compiler bug, an animated-WebP decoder bug and a Crashpad memory-safety bug to escape the renderer sandbox, which earned Google's first $250,000 full-chain reward under its April 2026 VRP rules, plus lessons for LLM-era vulnerability research.
17 min · 3,973 words
How we built the AssetHoard MCP server
AssetHoard’s Mark Gandolfo explains how the library’s built-in MCP server lets Claude Code, Cursor, and other agents search owned game assets and export into a project—and why the server is locked down so file contents stay on the user’s machine.
10 min · 2,254 words
One-shotting an Unreal Engine plugin from a 3,000-line plan
Jack Minnetian describes writing a 3,000-line plan for a voxel painting tool in Unreal Engine, handing it to an agent that returned more than 25,000 lines in fourteen hours, and what the success and failure modes say about plan-driven coding agents.
15 min · 3,353 words
Redesigning source control at Anthropic
Anthropic explains how merges to its monorepo grew from about 2,500 to over 30,000 per day in 2026, why git hosting became the bottleneck, and how it built and migrated to Scion, an in-house version control system, in five months, plus lessons and ideas for the wider git ecosystem.
16 min · 3,745 words
AI models may be capable of self-replicating without access to a datacenter
Anthropic's Frontier Red Team finds small open-weight models may be months from self-replicating across cheap IoT devices such as off-brand Android TV boxes: a 9B model set up inference on 8 of 9 hardware types, and GLM-5.2 completed end-to-end replication in simplified evals.
23 min · 5,391 words
Three weeks of agentic refactoring: could we trust the result?
A case study of three weeks refactoring DATAMIMIC CE with coding agents: how a before/after verification harness was silently testing new code via an editable-install finder, which results were voided, the architecture guardrails used, and what the team will change next.
6 min · 1,375 words
Why aren't AI companies being charged for hacking?
Edaqa Mortoray argues that recent unauthorized-access incidents by AI agents from Anthropic, OpenAI, Google and Meta are not unpredictable accidents: they follow foreseeably from deliberate design choices, which establishes the intent that hacking laws require.
10 min · 2,289 words
Recursive Self-Improvement through Multi-Agent Self-Supervision (MASS)
Sakana AI and collaborators present MASS, which repeatedly optimizes multi-agent workflows and distills their collective intelligence into a self-supervision signal for recursive self-improvement, improving score per output token over the base model on research benchmarks.
15 min · 3,501 words
Brooks Lybrand asks whether web frameworks still matter in the age of vibe coding and agents—and whether there is still room to build new ones after React’s win.
19 min · 4,419 words
Anomaly’s OpenCode 2.0 overhaul: one local background service shared by TUI, web, and desktop; Code Mode for MCP tools; background subagents; hot-reloadable plugins; and an embeddable @opencode/sdk with breaking V1 plugin/API changes.
4 min · 809 words
SlashDocs: the platform behind 28k+ documents at Razorpay, now open source
Razorpay open-sources SlashDocs, the self-hosted MIT-licensed hub where AI agents publish interactive HTML docs and humans review them with anchored comments, versions, CLI, and MCP—powering 28k+ internal documents.
9 min · 2,045 words
Meta AI Research introduces the Muse Wonder series (Muse Wonder and Muse Wonder Full), frontier models aimed at personal and professional agentic tasks that combine Active Memory and Active Focus, available through the Meta Model API and Muse Code.
4 min · 1,030 words
"200 OK" doesn't mean your AI agent did the job
Prathap K argues that a success response from an AI agent's tool call only means a request was accepted, not that the work happened, and describes the read-back check he uses to verify each write against the real system (email, CRM, messages).
2 min · 562 words
Two Hills: Weights and Harnesses
Part 1 of an LLM agent evaluation series: when a model system is not good enough, you face two hills—improving model weights versus improving the harness around the model—and how to choose where to invest.
6 min · 1,328 words
Unikernels were hard. Key word: were.
Geoffrey Huntley writes up his conversation with Justin Cormack about the unikernel revival: why MirageOS-era unikernels were hard, the operating system as design debt, porting missing libraries with AI agents, Nix machine tests, security trade-offs, and languages designed for agents, with chapter links to the recording.
13 min · 2,922 words