Mistral Agents API(mistral.ai) |
Mistral Agents API(mistral.ai) |
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.
I was looking around at Le Chat, a thing I haven't done in months, and I thought that they've really worked on interesting stuff in interesting ways.
The ability to enrich either a chat or generally an agent with one or more libraries has been solved in a very friendly way. I don't think OpenAI nor Anthropic have solved it so well.
It certainly looks easy to implement, I will say that! Docs halfway down the page: https://docs.mistral.ai/agents/mcp/
Mostly this seems like an end-run around tool calling scalability limits. Model performance degrades heavily if the field of possible tools gets too large, so you insert a component into the system that figures out what tools should be in-scope, and make only those available, to get reliability higher.
In terms of "why outsource this" it seems like the idea is that their orchestration agent would be better than a cruder task state machine that you would implement yourself. Time will tell if this assertion is true!
Where do you see that? That would be neat, I'm under the impression orchestration is manual though – you define an agent and give it the ability to hand off tasks to sub-agents.
1) It's really hard to follow some of the videos since you're just copy pasting the prompts fr your agents into the chat because the output generation comes out and hides the prompts. Instead put the prompt text as an overlay/subtitle-like so we know what you're doing
2) The clicking sound of you copy pasting and typing is not ASMR, please just mute it next time
3) Please zoom into the text more, not everyone has 20/20 super vision 4K style
But even at maximum 1080p resolution, the image quality is not that great. And while we are at it, the wine-red (#833048) on dark-brown (#23231F) syntax highlighting for keyword arguments has very poor contrast ratio of around 1.8 to 1: https://webaim.org/resources/contrastchecker/ which earns a rating of "Fail" across the categories normal text, large text and UI elements.
Very sloppy job, imo.
It costs next to nothing to come up with a little story and have someone on Fiverr narrate it (or an AI, after all that's what they sell).
Ultimately these are product/service companies, levering their research and innovations as differentiators.
If you're "only a model" company you likely have no moat.
If there's a local European option that does most of what an American or Chinese company does, that's simply a safer choice.
From this point of view, them trying to do everything at once makes a lot of sense. They don't actually need to be the absolute best or even the cheapest at any one thing. They need to just exist in Europe, be stable, and offer good services that people want. Casting a wide net is a better strategy for them.
The pitch is that if you do this bucketization, the overall orchestrator can intelligently pick the bucket to use, but the idea is that at any moment the LLM is only exposed to a limited set of tools.
As opposed to the more pie-in-the-sky idea that given N tools (where N is very very large) the LLM can still accurately tool-select without any developer intervention. This seems pretty far off at this point.