Heso-AI-Orchestrator
A two-agent system that automatically reviews and improves your codebase in a continuous loop, one agent thinks and plans, the other does the work, sending you Telegram updates without needing to babysit a terminal.
HESO is a system for running two AI agents together in a continuous loop to automatically review and improve a software codebase without human involvement in each cycle. The name stands for Hermes, Sisyphus, and Orchestrator, referring to the two agents and the overall coordination layer.
The core idea is a strict separation of roles. Hermes acts as the thinking half: it reviews the codebase, checks whether the previous task was completed correctly, generates ideas for improvements, and decides what to work on next. Sisyphus acts as the working half: it receives a single clearly defined task from Hermes and executes it, then reports back. Sisyphus never decides what to do, never reviews its own output, and never sets priorities. The README argues that most automated code review loops fail because a single agent reviews its own work and reliably finds nothing wrong. Keeping the reviewer and the worker separate is the central design choice.
The loop runs on a schedule, using standard cron and the tmux terminal multiplexer rather than any specialized infrastructure. There is no vector database or machine learning model running locally; all state is stored in a plain SQLite file. The README states the design targets 800 to 1,000 cycles per day on ordinary hardware. Each cycle ends with a short plain-language message sent to you via Telegram, so you can monitor progress without watching a terminal.
The system has three memory layers: a file that holds your stated intention for the project (written once and not changed without asking you), a set of database tables tracking ideas, user experience notes, and pain points, and a keyword pool used to introduce variety into the brainstorming step. The full database schema, prompt templates, and task format are documented in a SPEC.md file in the repository.
The project is in alpha. It is released under the MIT License.
Where it fits
- Set up a fully automated coding loop that reviews and improves your project overnight with no manual intervention.
- Monitor an AI-driven code improvement agent via Telegram messages without watching a terminal.
- Run continuous AI-assisted refactoring on a codebase using only ordinary hardware and a SQLite file, no GPU or vector database needed.
- Use the reviewer-worker separation pattern to avoid the common failure mode where a single AI agent reviews its own output and misses everything.