Blog

The agent writes the film as code, headless Chromium captures each frame, FFmpeg encodes. LaunchVideo spends 100k tokens a film; a 14-agent reel cost $30.46.

• coding-agents
• claude-code
• video
• ffmpeg

A durable harness commits every step of an agent run to storage, so a crash resumes mid-turn. What Pi Durable shipped on 1 October, and what is only claimed.

• harnesses
• coding-agents
• ai-agents
• durable-execution

GitHub was the most added tool on Stackness in September 2026. Supabase and Cloudflare went from no profiles to 6 and 5. The table and an npm check.

• monthly-roundup
• trends
• ai-coding-tools

Sonnet 5.5 costs half of Opus 5.5 per token, but cache reads cost $0.20 on both. Which model plans, which one edits, and what one metered run saves.

• claude-code
• coding-agents
• model-routing
• benchmarks

A frozen eval panel is the only rerunnable way to tell whether a hosted model got worse. What livenerf measures on Opus 5.5, and how to size your own.

• benchmarks
• evals
• claude-code
• coding-agents

You can get Jev-style typed decisions from an ordinary LLM by forcing one output token and reading its logprobs. Where that matches Jev, and where it loses.

• system-one-models
• jev
• calibration
• logprobs

An ADE is an app built around tasks that coding agents run, not files you edit. What JetBrains Air shipped, and whether an ADE replaces your IDE.

• coding-agents
• harnesses
• ai-coding-tools
• ides

Facts every session needs go in AGENTS.md, procedures go in skills, and rules that must hold go in hooks. How to share one handbook across six agents.

• agents-md
• claude-code
• context-engineering
• agent-skills

Opus 5.5 sends its between-tool notes as empty thinking blocks and ends some turns early. What breaks in an agent harness, and the documented fixes.

• claude-code
• coding-agents
• harnesses
• prompt-engineering

To run a System One model locally you pick a runtime as well as a model. Ollaya serves ten open decision models behind a Jev-compatible API, laya-mlx runs Laya on Apple silicon, and llama.cpp needs a shim. Which numbers are measured, and when hosted Jev still wins.

• system-one-models
• jev
• local-ai
• benchmarks

Weekly limits tightened and GPT-6 Luna landed at half the price of GPT-5.6 Luna in the same week. The plan and price numbers developers measured in September 2026, and the planner plus executor split they point to.

• coding-agents
• ai-coding-tools
• model-routing
• claude-code

A Claude Code mod is a plugin whose behaviour is a TypeScript hooks module running inside Claude Code's process. What shipped in 2.1.277, what is only announced, and which extension point each change belongs in.

• claude-code
• coding-agents
• harnesses
• plugins

JevBench scores Jev-class decision models on four axes: accuracy above chance, calibration, speed and cost per 1,000 decisions. Its v1.4.1 board puts Jev first at 63.3. A day later, a post argued that no benchmark can make these probabilities trustworthy on your data. Both are right, and this post covers what to measure yourself.

• system-one-models
• jev
• benchmarks
• calibration

Since 2.1.277, Claude Code reads AGENTS.md only when no CLAUDE.md or CLAUDE.local.md sits in the working directory or above it. Until 2.1.281, a remote feature flag also skipped the file without a warning for anyone with telemetry off. Here is the precedence table, how to check what loaded, and which other agents read the same file.

• claude-code
• agents-md
• coding-agents
• context-engineering

A harness is everything in a coding agent except the model. Two September 2026 studies measured which parts of it matter: context management when the window is tight, planning mostly as a cost saver, and the harness as a whole far more for your bill than for your pass rate.

• coding-agents
• harnesses
• claude-code
• context-engineering

A week after launch, Jev had a trusted monitor, two computer-use agents, three self-hosted clones, a benchmark and a chatbot nobody should use. Here is the stack slot each project fills, and which of their numbers anyone has shown.

• system-one-models
• jev
• ai-agents
• benchmarks

Compaction is the step where a coding agent summarises its own history to keep working. Here is what it keeps, what it drops, what OpenAI found hiding in its summaries in September 2026, and the moves people use to compact less.

• coding-agents
• context-engineering
• claude-code
• compaction

The open-source alternatives to Jev are Laya, Kev, SemIf, Open-Jev and Nimble. They are four kinds of project, and here is what each one replaces.

• system-one-models
• jev
• open-source
• local-ai

A System One model answers typed questions with typed answers and a probability, not text. It sits beside your LLM and takes the small, frequent decisions.

• system-one-models
• jev
• llsm

The first edition of our AI coding tool switch series starts with a very small September 2026 sample. Cursor and Claude Code lead, while co-use is more common than switching.

• ai-coding-tools
• coding-agents
• cursor
• claude-code