Skip to Content
AIAgent Autoloop Study

Study: Agent Autoloop

Published on

Agent Autoloop  is a self-hosted project for coordinating command-line agents such as Claude Code and Codex. Its English README  describes a loop that accepts an event, chooses an agent and working directory by rule, runs one round, and records the result.

When the agent needs a decision, the loop sends a card and resumes after a person answers. This page is a study of that design, based on the project’s documentation; it is not an independent test of the software.

Trace one piece of work

  1. Event: A schedule, issue, or alert enters through an event source.
  2. Rule: Configuration selects the agent, task, and working directory.
  3. Round: The agent runs and places its result in its final reply.
  4. Decision or finish: A completed task ends; a question becomes a card, and the agent process exits while waiting.
  5. Resume: A human answer starts the next round of the same conversation.

This flow separates the long-lived coordination state from the short-lived agent process. The README’s workflow section  is the source for these steps.

Design choices to examine

  • Durable state: The project says it stores an append-only event log and indexes in SQLite. Ask which writes must be committed before a caller receives success, and how the loop reconstructs state after a restart.
  • Plugin boundaries: Event sources, agent adapters, notification channels, and delivery actions are separate extensions. Trace which component owns credentials and which component decides what an agent may do.
  • Human decisions: A card represents a choice that should wait for a person. Look at how the project records the question, avoids running the agent while waiting, and resumes the same conversation after an answer.
  • Loop control: The documented hop limit stops chains of agent-emitted events from continuing indefinitely. Examine what gets recorded when that limit is reached.
  • External input: The README describes a separation between untrusted event data and the rule that selects an agent and working directory. Check where that boundary is enforced in the code.

These are claims and design descriptions from the upstream README , not measurements made for this site.

A small study exercise

Read the installation and try walkthrough before running anything. In a disposable local workspace, observe one ask-and-answer cycle: note the event, the chosen rule, the agent’s final reply, the waiting card, and the resumed result.

Then use the documented agent-autoloop why command to trace the round and compare the trace with your notes. For a new rule, validate its configuration and use the documented dry run before allowing it to dispatch real work.

The project says some webhook templates were tested against local mock services rather than live providers. Keep that distinction in mind when evaluating an integration. See the plugin and testing notes .

Questions worth answering

  • What state survives a process crash, and what evidence shows that no acknowledged event was lost?
  • How are duplicate events and repeated answers handled?
  • Which actions require a person, and where is that policy defined?
  • What changes when the agent adapter is Claude Code, Codex, or a generic CLI command?
  • How would you test a failed notification channel without running the agent’s work twice?
Last updated on
Codex