Mastra (YC W25) is launching on Product Hunt today, releasing stable 1.0 for building AI apps and agents, and announcing 300,000+ weekly npm downloads, 19,400+ GitHub stars, and production use at companies like @Replit and @WorkOS.
I had a blast working on this launch with @calcsam @smthomas3 @bookercodes and team, and wanted to share with you some thoughts we put into the ops.
one thing i would want to understand better is how memory behaves when the same agent is serving many different users.
Mastra
@alira_salu I think a helpful distinction here is memory vs. knowledge. Mastra Factory is always building the team's knowledge base, recording facts relating to certain areas of the code or the process. Future agent runs, regardless of user, can reference this.
Does it retry automatically or does it require manual error handling everywhere?
Mastra
@john_michael31 Agents should recover the best they can. Naturally, some errors will require human intervention. We should also stop trying if it's futile, otherwise it would be a waste of tokens
Mastra
@john_michael31 you can enable automations and it will mostly work autonomously. However, we typically recommend starting on manual mode so you can build trust in the system first.
Mastra
Very excited to launch Mastra Factory! It's changed how our team ships code. My favorite parts of the factory:
1. It's more collaborative - it gets coding agent sessions off our local machines and into a shared environment. I can see how others on the team are steering their agents and learn from them
2. Tagging factory in slack makes it easy to ship one off tasks that would have otherwise have been forgotten
Mastra
@smthomas3 It's come a long way since we shared an early preview back at TS AI Conference in London back in July!
Mastra
@smthomas3 @bookercodes i also enjoyed Abhi's talk at Agent Night by @WorkOS (47:26)
Mastra
@smthomas3 @fmerian We should link this on the homepage lol. Such a good talk
Mastra
@smthomas3 @bookercodes cc @patrycja2 @paulscanlon 👀
Mastra
@smthomas3 S/O for this new launch! humbled to contribute to it again
Mastra
Factory has been a game-changer for our team. It's writing 30% of our PRs and closing 60% of our issues. Excited to share with the world!
Mastra
@sam_bhagwat I only see that number going up!
Mastra
@hamza_afzal_butt you are able to define a "triage" skill so you can determine what level of detail is needed before work is started. You can also configure the factory to require human approval at each stage if you want to monitor things more closely.
I would recommend manual mode in the beginning until you have a higher level of confidence in the work moving through the factory.
Mastra
@fmerian @hamza_afzal_butt Each step is an agent with a skill. In this case, the skill handles triage, and you can define how you want it to behave. As with any agent, you should test it thoroughly. One cool thing about Mastra is that you can set up tracing and evals to help with that. I can see people using online evals to keep iterating on their skills. We'll add some guides on this at some point.
Mastra
yes, and to learn more, read the docs here: factory.mastra.ai
I would probably start with a small workflow before trying to build a full agent system with it.
Mastra
@jabari_zuri That's a pretty normal reaction! I get it. You'll see the biggest upside when you embrace the system
Mastra
as suggested by @smthomas3 in this comment, you can start by running each step on manual mode until you have a higher level of confidence in the work moving through the factory. enjoy!
Mastra
@jabari_zuri we recommend starting small. Maybe just try it with smaller issues first and see how it does. Eventually, you can expand it as you trust it to do more complex work.
Mastra Factory is built with all Mastra primitives and is now used to build Mastra itself. We love doing the meta thing! Hope you enjoy the productivity gain like we have!
Mastra
@suhdudeitsabhi So meta!
Mastra
@suhdudeitsabhi spot on!
picture this: @Mastra is an active open-source project, with contributions from the community and maintenance by a core team, 100+ opened issues/week, and 200-250 merged pull requests. the factory automated 25-35% of the PRs and closed 50-60% of the issues in 3 months.
really looking forward to seeing what the community here will build with it.