Building @Mastra - teaching you how to build agents along the way
Joined July 2015
- Tweets 6,921
- Following 574
- Followers 9,084
- Likes 14,380
919 Photos and videos
Pinned Tweet
Jun 10
What's the difference between an eval and a test?
Yeah, I wasn’t totally sure either...
So I went deep and made Evals 101 - a short 20 minute mini-course on the fundamentals (with examples)
Fair warning - evals are hard. There are lots of ideas, opinions, and best practices floating around - TBH everyone is still figuring it out. No matter where you land on the right way to do evals, this mini-course gives you the foundation you need:
- What an eval actually is
- How evals differ from traditional tests
- What a scorer is, and how they work
- Code-based scorers vs. LLM-as-judge
- Online vs. offline evals
- How to start thinking about evals in real applications
4
4
30
1,798
Alex Booker retweeted
Jun 9
you can use Fable 5 with @mastra today
npm create mastra@latest
6
17
2,451
Alex Booker retweeted
Jun 9
agents are meant to be multiplayer
agents are meant to be steerable
agents are meant to be proactive
Signals makes this easy
Jun 9
Signals are a fundamentally new way to give your agent new input or context without restarting the agent loop.
They let you steer a running agent with a new user message mid-execution instead of stopping everything and starting again from the top.
But they’re not only useful while an agent is actively running.
A signal can arrive while an agent is stopped too.
For example, a coding agent might get a GitHub event, save that context, and either pick it up on the next run or wake up to handle it. We implemented this in our coding agent Mastra Code 😉
We also introduced `subscribeToThread`, which lets a client listen to an agent’s events without necessarily sending a message.
That means multiple clients can observe the same agent thread at once - whether that’s for multiplayer agent experiences or horizontally scaled Mastra instances coordinating around long-running tasks.
2
4
671
Jun 9
Signals are a fundamentally new way to give your agent new input or context without restarting the agent loop.
They let you steer a running agent with a new user message mid-execution instead of stopping everything and starting again from the top.
But they’re not only useful while an agent is actively running.
A signal can arrive while an agent is stopped too.
For example, a coding agent might get a GitHub event, save that context, and either pick it up on the next run or wake up to handle it. We implemented this in our coding agent Mastra Code 😉
We also introduced `subscribeToThread`, which lets a client listen to an agent’s events without necessarily sending a message.
That means multiple clients can observe the same agent thread at once - whether that’s for multiplayer agent experiences or horizontally scaled Mastra instances coordinating around long-running tasks.
4
6
46
6,124
Jun 5
My team at Mastra just quietly invented agent signals
- Subscribe to a long-running agent from anywhere
- Watch the same run from desktop, mobile, or Slack
- Let multiple people steer it without stopping the loop
Your agent can subscribe to system events too - like a coding agent getting notified that a PR has a new comment, then changing course instead of wasting time or tokens
P.S. Signals is in alpha and so is my demo 😅 Excuse the bug at the end. This is too cool not to share!
16
10
151
16,171
Jun 5
That moment of anticipation when you’ve come up with the perfect plan with your coding agent.
Now it's time to build and see…
2
333
Jun 4
Since we started Mastra, people have asked for an agent builder. Today we’re launching one!! 🚀
Not a generic builder... but one designed for engineer-led teams.
Developers define the tools, models, and guardrails in code - everyone else uses a chat-based builder to describe the agent they want, test it, and build on top of the primitives engineering exposes.
That means PMs, ops, support, sales, and other teammates can prototype useful agents without waiting on the backlog, while engineering still controls auth, permissions, allowed tools, storage, model defaults, and cost.
Anyone can publish to an internal agent library or deploy their agent to Slack with one click (Teams coming soon!)
And should an agent outgrow the builder, you can eject it into a TypeScript project and graduate it to an engineering project. This is very unique!
6
8
48
6,819
Jun 1
The Mastra server is a fundamentally Hono server
So you can add middleware
Here's a handy one I've been using lately to log the current user's role
Handy when debugging auth
1
1
10
851
Alex Booker retweeted
May 29
We're launching Agent Builder, your internal agent platform:
- Devs write tools, select models, build workflows
- Non-dev teammates assemble into agents (or vibe code), chat, publish
- Includes RBAC, ownership, and visibility controls
- Under the hood everything's Mastra agents
29
39
467
110,694
New podcast, new format. Three founders join us.
Waste Tokens, Save Time
00:00 Three Frontier Founders
01:27 AI Software Factories
04:15 Waste Tokens, Save Time
05:47 Models Instructing Humans
09:30 Is Pure Software Dead?
12:04 You Don't Get Stuck Anymore
With @rauchg, @maxhodak_, and @bscholl.
147
299
3,439
269,037
May 28
In Mastra too 😉
May 28
if you havent yet announced that opus 4.8 is available* in your product youre unfortunately ngmi
it was released 1hr ago, what are you even doing
*available in raycast
4
498
May 28
🔴 Live with @wardpeet, Mastra head of open source, teaching you all about A2A and ACP x.com/i/broadcasts/1mxPaaWMo…
2
5
306