FailMore parent
Is this basically a LLM that has tools automatically configured so I don’t have to handle that myself? Or am I not understanding it correctly? As in do I just make standard requests , but the LLM does more work than normal before sending me a response? Or I get the response to every step?
The aspirational goal is that the model knows what tools to call and when, without human intervention. In practice, you'll see varying efficacy with that depending on the tools you need. Some of the tool usage is in-distribution / well represented in training set, but if you have some custom exotic MCP server you created yourself (or pulled off of some random github) you may see mixed results. Sometimes that can be fixed by simply augmenting your prompt with contrastive examples of how to use or not use the tool.
As an aside, my experience with devstral (both via API and locally w/ open weights) has been very underwhelming to this effect. So I'm curious how this new agent infra performs given that observation.
It's a software framework for orchestrating agents. Each agent can have its own system prompt, its own tools, and it can delegate ("hand off") to a different agent. When a hand off occurs, the LLM runs again but as a different agent.
Like Gemini Gems, but agentic?
Gemini Gems seems to be a ChatGPT “GPTs” equivalent, and I never figured out what those actually are. Mistral Agents API is like OpenAI Agents SDK.
Gems and GPTs are just a way to customize the system prompt from the web UI.