Closed Adefioye closed 3 weeks ago
Tested with gpt-4o-mini for any regressions and with anthropic 3.5 haiku. Works great. with both The models behave slightly differently and anthropic seems to many times go beyond what is expected in the task (e.g. try to check out from amazon when the task is add product to cart etc). It suggests it might benefit from a custom prompting.
If grok works as expected, i think this PR can be merged. Thanks @Adefioye !
@Adefioye this seems to work all around and does not break original Agent-E thank you. I will merge it.
One thing to note, making this work anthropic LLMs would be most advantageous for project adoption. I think if you want, you can fiddle with the prompts for both planner and browser nav till they work with something like Haiku. Once they do, you can commit that prompt in the example file agents_llm_config-example.json
(adding it to the entries that you made). With groq I kept getting rate limited, so I tested as much as I could, but it shows good promise.
As a developer, I want to be able to use anthropic and groq-based models to experiment with Agent-E.
agents_llm_config-example.json
.run.sh
file to runpython -m -u ae.main