Agent-Autoloop is an open source tool for Windows, macOS and Linux, published under the MIT licence and written in TypeScript. The project was started on 3 November 2026 and reached four thousand stars within a week. Its target is narrow and practical. A command line agent is capable enough today, but somebody has to sit in front of the terminal, answer one question at a time, and the run dies the moment that person walks away. This project positions itself as the scheduling layer in between, so the agent picks up work in the background and comes back only when something needs a decision.

Version 1.7.1 arrived on 9 November 2026.

What the layer absorbs

Between arrival and delivery a task passes four handovers: where the work comes from, which agent gets it, who answers when it stalls, and how the result ships. Agent-Autoloop turns each of these into a plugin. Event sources connect to GitHub or GitLab hooks, an HTTP endpoint or a timer. The agent side wraps a command line tool. Channels push to a desktop notification or to a chat service such as Feishu, DingTalk or WeCom.

Plugins mean a single stage can be swapped without reinstalling the whole set. Change the agent in a configuration line and the rest keeps running, with no waiting for official support. The project also states plainly that there is no dashboard. When a decision is needed it pushes a card, and the global picture comes from one command.

A waiting run exits instead of idling

This is the cheapest idea in the design. When an agent asks a question, many tools keep the process alive until somebody answers. An hour of waiting holds an hour of memory, and a token-billed path keeps counting in the background. Agent-Autoloop exits immediately and writes the waiting state into a database. Once the answer arrives, work resumes from the same point in the conversation.

The documentation calls this a suspend with no consumption, and it is explicit: three days without an answer means three days with no agent process running. For a single user that saves a little power. For anyone running dozens of rounds a day, it shows up on the bill.

Outside text is data, not instruction

The second idea worth noting is the separation of instruction and data. Event bodies usually come from outside, for instance the title and description of an issue, and that text may well contain a line telling the reader to ignore previous instructions. Agent-Autoloop never splices that text into the prompt. It stores it as a file, and the prompt carries only a validated type, number and link.

The same reasoning governs permissions. Which agent runs, in which directory and with what rights is decided by local rules alone. An event that carries a working directory, an agent name or a permission field is rejected outright.

Event sourcing and a depth limit

Two further mechanisms are plain engineering. Event sourcing derives all state from an append-only log, so a killed process loses nothing that was already confirmed, and one command explains why any round ran. The depth fuse borrows the idea of TTL from networking: an event handed back to an agent increments a counter, and past 8 levels the program only records it and pushes a card instead of dispatching more work.

Three releases in a week

The pace is quick. Version 1.6.0 on 6 November 2026 added native Windows support. Version 1.6.1 right after it fixed an installation failure caused by Chinese user names and Chinese directories, along with a leftover process window on the desktop. Version 1.7.1 on 9 November 2026 repaired a macOS service that refused to start: a terminal launched from the graphical session wrote a configuration file that everyone could write to, so the service loader rejected it. Every installation now resets permissions to 644, and one reinstall rescues a machine that is already broken.

Version 1.7.0, released the same day, fixed a detail that is easy to overlook. The answer card now carries the selected text, where previously only an index appeared, and an agent would read the second item as the first.

What the authors admit is untested

Webhook templates

Four templates were verified only against a local mock service.

GitLab path

Tested against the documented format, never against a real instance.

Third party agents

Connection commands were assembled from documentation and not run one by one.

That habit of marking the untested parts is more credible than a page of claims about full support. What matters in Agent-Autoloop 1.7.1 is not how many agents it can wrap but that it drives the cost of waiting almost to zero.