artlu's Bear Blog

Mastra initial impressions

Heartbeat-based agent orchestrators work properly 9 out of 10 times, or maybe 99 out of 100 times. But based on my own experience as well as things people share online, the schedulers tend to fall over, and people spend many cycles trying to stand them back up.

A crontab runs properly 9,999+ out of 10,000 times. It can be slightly harder to reason about how to get information where it needs to get to, but the reliability is well understood.

One would still prefer to build scheduled tasks in a 'Claw-style orchestrator when:


Mastra is to smithers, as Hermes is to pi

Mastra is an Apache 2.0-licensed project founded in 2024 (pre-OpenClaw!), with over 14k commits

It has a freemium business model

It ships a "Build with AI" project skill and a docs mcp

Starter template is easy to get working

Never gives the "magic" feel of OpenClaw or variants. Dev is always under control

In about two hours, I migrated all my scheduled jobs from Spacebot1. They now live in Mastra Workflows, or deterministic python scripts on a system crontab. They play nicely with each other as distinct users in shared channels in a private Discord server


Expectations

with OpenClaw, you can just ask it to do anything and it will basically figure out how to make it happen. It will probably require several handholding iterations before appearing to reach a satisfactory state. Then it will probably fall over 5-20% of the time, depending on a mixture of your skill and your model's capabilities

with Mastra, I expect to think harder upfront, with a tighter aperture of how to explore possibilities. But I expect after something has been scoped down so that ambition matches capabilities, it will continue to work

Less magical, more limited, more robust


A Brief Detour

I also tried aeon before Mastra. It looks even lighter and more flexible, like the light frameworks I tend to prefer

but to get light, it is much more opinionated and those opinions don't match mine lol

it also gives new users a cluttered experience, which I understand because it wants to highlight capabilities

but I want to swap everything for things that give me more control, and it felt clear I'm not the target user. Aeon's ideal user is someone who is willing to pay providers for great defaults. I'm more on the DIY side and look forward to swapping everything for my own shi



  1. Spacebot operates most of its scheduling in reliable Rust, but relies on agentic delegators to spawn workers and figure out the appropriate level of context to pass to children