This week's AI automation coverage centred on agents doing work that used to need whole teams. The most concrete case is Cockroach Labs, which says a pipeline of coding agents added IBM Db2 migration support in under two days. A developer also used the ML-Intern agent to build seven custom models in days.
The rest of the set leans toward tooling around agents: visibility into multi-agent sessions, specs checked before code is written, a macOS pointer for on-screen steps, and a project that takes agent-built apps live. Several pieces also carry caveats, from a preliminary archive search to a penetration testing repository with strict usage limits.
1. Building Seven AI Models With Hugging Face's ML-Intern Agent
A developer on the Hugging Face blog describes building seven custom AI models in days with the ML-Intern agent inside HuggingChat. Each project began as one chat message and ended as a public Hub model with an evaluation. The detail that matters: the agent asks for a budget before spending and runs a small test first. Costs ranged from a few dollars to about USD 37.
Best for: Teams testing agent-driven model training on a small budget
2. AI Coding Agents Add IBM Db2 Migration Support in Under Two Days
Cockroach Labs says a pipeline of AI coding agents added IBM Db2 source support to its MOLT migration tools in under two days, with no human writing the code. The company compares it with the Oracle work in 2024, which took nine months and roughly $160,000 in engineering time. The Db2 token bill came to $4,172. Useful for database teams weighing agent-built migration tooling.
Best for: Database teams evaluating agent-built migration tooling
3. AI Agents Decompiled a Shooter: Lessons From 500+ Billion Tokens
A developer on Maurice's Blog describes three months spent using autonomous AI agents to decompile a popular first-person shooter, aiming for an accurate, stable and feature-complete rebuild. The estimated 600 to 700 billion tokens consumed frame the cost, while the workflow lessons apply beyond games. The author has not named the game, and two earlier posts were removed. Byte-level verification is the key detail for engineers.
Best for: Engineers interested in agent workflows and verification-heavy reverse engineering
4. Reverse Engineering with an AI Coding Agent: How REA Works
REA is a reverse engineering toolkit for AI coding agents, presented on a public website with setup instructions, guides and interactive examples. An agent inspects a running or compiled program, explains its behaviour, then changes or rebuilds it. Examples include a browser dinosaur game and the percent button in Windows Calculator. The page was published on 11 October 2026.
Best for: Developers curious how agents can analyse existing software
5. Agent Standup Turns Claude Code Sessions Into a Live Team Call
Agent Standup is a free, MIT-licensed open-source tool that shows Claude Code multi-agent sessions as a team video call. Each agent appears as a participant, showing which is progressing, which is waiting on another, and which needs a human decision. It reads session files on your own machine and its makers say it adds no extra token cost.
Best for: Developers running several Claude Code agents who want a status overview
6. Intent-Router Turns Vague AI Requests into Checked Specs
Intent-Router is a GitHub project described as an intent compiler. It turns vague requests, such as adding caching to a user API, into typed specifications before any code is written. It searches available files for answers and asks the human only about questions no file can settle. Its README describes three passes, starting with parse. Useful for teams that want specs checked before agents write code.
Best for: Teams wanting agents to check the repo before asking questions
7. GoLive Agent Skill Aims to Take Agent-Built Apps Live
GoLive is an open-source Agent Skill and Node.js CLI that takes agent-built apps live on hosting, database, domain, email and payment accounts you already own. The project is in early alpha, at version 0.1.0-alpha.8. It asks for approval before every change, which matters for anyone deploying their own work.
Best for: Solo developers deploying agent-built apps on their own accounts
8. bigarrow: Free macOS Tool Lets AI Agents Point at Your Screen
Bigarrow is an MIT-licensed macOS command-line tool that draws a large arrow and text sign over any window, so AI agents can show you where to click. Clicks pass through the arrow, keyboard focus stays put, and the arrow removes itself when done. It is packaged as a skill for Claude Code and Codex. Useful for anyone guiding agents through on-screen steps.
Best for: Developers whose agents need to point at on-screen elements
9. ARTEX: Open-Source AI Agent Built for Automated Penetration Testing
ARTEX is an autonomous penetration testing system built around LLM agents, and a GitHub repository now keeps it as a final source-only backup. The README describes it as the champion entry in Baidu's agent+ attack-defense challenge. The original Docker image has been removed from its earlier location, and distribution has moved to the maintainers' own accounts. Read the usage limits before using it.
Best for: Security researchers studying agent-based penetration testing tools
10. AI Historical Archives Search Uncovers Forgotten Meteorite
A software engineer used AI tools to search Dutch East India Company records and old newspapers, reporting a forgotten meteorite fall and three rhinoceroses that never reached their recipient. The account is preliminary, and each finding still needs specialist review. It follows a historian's dodo post from October 1, 2026. A clear example of machine search reopening unread archives.
Best for: Researchers and archive readers interested in AI-assisted search
The takeaway
Practical takeaway: before handing a task to an agent, write down the budget, the test and the approval step. Several of these projects, from ML-Intern to GoLive, depend on that discipline, and it is the safest place to start.
Comments (0)