Learn how to build powerful AI agents with Mastra. Master tools, memory, and MCP integration to create sophisticated AI applications.
somehow i hadn’t heard of mastra given that i’ve gone out of my way to find and try basically every framework in this vein under the sun, but i gotta say: i’m quite impressed in terms of how clean it is and how well it aligns with i would consider “real” work in this arena… so many of these frameworks try way too hard to be batteries included and/or are otherwise so simple / limited as to feel like they were made by people who don’t actually… “get” it and thus that by extension its users don’t either…
I've been using Mastra for a few months now and we've really been liking it. Team is super responsive and a lot of improvements are afoot.
thanks for checking mastra out
if you do find things that need improvement, please let us know in our discord or on x!
will do!
Newbie here: how much of creating agents are portable from one AI platform to another? Do they mostly have the same concepts but differ slightly in implementation? Or something else?
The concepts are pretty similar but this course is definitely pretty Mastra specific.
If you are looking for a more general overview on agent development, my co-founder wrote a short book (available free at https://mastra.ai/book)
Most agent frameworks implement a ReAct (Reflect->Act) pattern in the form of a loop that "reasons" until a "final_answer" has been found. Most framworks also implement some form of session cache (scratchpad), tool use, and thought logging/tracing.
While most implementations are similar they are not easily transferable.
MCP is portable but isn't very useful for creating agent workflows.
This was interesting, but continues to shatter the illusion for me that LLM based programming is vastly superior to someone not using an LLM. It consistently struggled for me to set up what it was trying to do, and had little context in helping me work through issues. I was using windsurf so maybe that was part of my issue? I gave up in part two when it got in a never ending loop trying to connect to the GitHub mcp.
We're going to make some recommendations for which models work best with the course! We've seen some different experiences based on model and IDE.