Hello, I'm the author of this. Sitelooper is an open-source browser automation CLI containing a mini-agent. You give it the steps of a test in natural language and it runs a cheap model the first time and every step it gets right is compiled into a procedure that can replay with no model (or fall back if it needs to). Then this can be turned into a plain Playwright spec. If the app changes, only the broken step needs the model.
The bigger idea here is 'progressive automation': automatically converting appropriate, more predictable parts of an agent's work to cheaper models then deterministic code, and reverting this if it breaks.
To say upfront: the benchmark is mine (all runs & misses published) and about 5% of runs give false passes. Repo: https://github.com/restspace/sitelooper. Any questions welcome.
Hello, I'm the author of this. Sitelooper is an open-source browser automation CLI containing a mini-agent. You give it the steps of a test in natural language and it runs a cheap model the first time and every step it gets right is compiled into a procedure that can replay with no model (or fall back if it needs to). Then this can be turned into a plain Playwright spec. If the app changes, only the broken step needs the model.
The bigger idea here is 'progressive automation': automatically converting appropriate, more predictable parts of an agent's work to cheaper models then deterministic code, and reverting this if it breaks.
To say upfront: the benchmark is mine (all runs & misses published) and about 5% of runs give false passes. Repo: https://github.com/restspace/sitelooper. Any questions welcome.