The Secret Agent! Blog
Straight from the codebase: how AgentiLoop Agent! works, what changed, and why. 100% Fresh.
Back on Hacker News, Six Months Later, with Auto-Pilot
Agent! is on Hacker News again as a Show HN. The April thread was about a native Mac coding harness; this one is about Auto-Pilot, the goal-driven loop that keeps running cycles until the goal is reached, and the Mario Kart-style game it has been building. Here is the post, what is behind each claim in it, and where to join the discussion.
GoKart: A Mario Kart-Style Racer That Auto-Pilot Built in an Afternoon
Give Agent!'s Auto-Pilot one goal - "create a Mario Kart clone called GoKart" - and come back to a Godot 4 racer with three tracks, eight items, AI rivals and 3,344 passing test checks. Two days and 87 agent commits later it is GoKart 0.0.2, with four courses, Battle mode, Time Trials and a Mario Kart 64 style menu. Here is what the log says actually happened, including the parts where it got stuck.
Agent! Gets a Face: An Optional Speaking, Animated Avatar on New Tabs
Coming in the next Agent! pre-release, scheduled for Saturday, October 10 - an optional avatar that speaks and animates alongside the task log. Not cute, not cuddly, and not a mascot. Now in development.
The Whole Family Ships: Agent! 1.1.87 and AgentiLoop CLI 0.0.5
Agent! for Mac gets Auto-Pilot, six new providers, a stricter critic and macOS 14.6 support. The Rust and Go CLIs get a bigger toolbox: search, web fetch, todos, AGENTS.md, /undo, custom commands and --json.
A Release on the First of Every Month, and Pre-Releases in Between
Agent! now ships a stable release on the 1st of every month. Daily pre-releases (moving to weekly) are where new features and fixes get tried out, and release candidates lock things down before the big day.
Sonoma, Intel, and the Mac That Was Not Dead Yet
Agent! for Mac now runs on macOS Sonoma 14.6 and later, on Apple Silicon and Intel. A lot of you asked for a pre-macOS 26 version. Here it is.
What Is an Agent Loop? A Robot, a Sandwich, and the Art of Trying Again
Look, choose, act, check. A playful, illustrated guide to agent loops—simple enough for a five-year-old, with plenty for the grown-ups.
Why Agent! Compacts at Half the Window, and the Three Bugs That Shrank It to 16K
How Agent!'s context compaction decides when to summarize a long task, and the September 28 fixes for fallback models, Ollama and a 272K Codex window.
Archive
Older posts that found few readers. They stay online and come back to the list above if they pick up again.
- How Agent! Got Started: Three Days in March
- The Terminal Strikes Back: Taking the Agent Loop Cross-Platform in Rust and Go
- Agent! Now Runs on macOS 14.6 and Intel Macs: How We Got Off macOS 26
- Before the Model Gets a Vote: Inside Agent!'s Shell Guardrails
- Welcome to the Agent! Blog: One App, Any AI, Total Command Over Your Mac
- Meet Jev: A Second Opinion That Answers in Types, Not Prose
- Why We Ripped Out Coding Mode: Let the Model Pick Its Own Tools
