- Author
- Arda Tasci
- Published
- Undated
- Kind
- Personal blog post presenting a design and a project
- Original
- eardatasci.github.io/c/ambiance/index.html
- Subject
- Harnesses and agent loops
What it is
Arda Tasci writes about how to free language models from the chat pane. The post lists what makes a good harness. It then sketches Ambiance, the author's own harness built on a Unix analogy.
What it says
- The post lists four goals for a harness. They are intuitive, transparent, lean and survivable.
- The author says a harness should reuse what models already know. Systems administration is common in training data.
- The post maps harness parts onto Unix paths. Tools go in /bin, logs in /var, skills in /usr/share/doc.
- The post says external data should be cleaned into plain text first.
- The author built an event bus called the Kernel. It watches files for changes and then invokes the model.
- The post contrasts the Kernel with OpenClaw's heartbeat. A heartbeat runs a full agent turn every 30 minutes by default.
- There are three default LLM users. They are root, pai and librarian.
What evidence it gives
The post offers argument and a design sketch. It gives a mapping table from harness parts to Unix directories. It shows no benchmarks, trials or user results.
Interests and limits
Arda Tasci promotes Ambiance and links an install script. The author therefore has a stake. The post says it is a work in progress. The post is undated. The author says Unix is a motivating analogy, not a direct comparison.