Michael Gabriel

AI Platform Engineer @ MaintainX @ Autodesk.
Shitposts about AI engineering and dev tools.
All opinions are my own.

[Michael Gabriel @mike_gabriel]

https://t.co/iXyMN8wCRL

[Michael Gabriel @mike_gabriel]

@gzcl3000 The Buddha literally threw away his nobility and money to become a homeless beggar, but pop off king

[Michael Gabriel @mike_gabriel]

@wgreklek @mdnlabs @bot Do you also use other XAI products like Cursor or grok build? I'm on the $30/mo plan and this is my usage after using it normally most of the week and purposely trying to burn through as much as possible today.

[Michael Gabriel @mike_gabriel]

It can be incredibly tempting to start from a finished Agentic product, but you can easily take something like PydanticAI or Mastra and set up your own infra around it.

Your DevEx should be based on your dev's needs, not what your SDK allows.

[Prasenjit Sarkar @stretchcloud]

Five agent SDKs now exist for roughly the same reason. Whoever defines how you build an agent keeps the developer relationship no matter which model ends up running underneath it. Cursor added itself to that list this morning, shipping an Agent SDK that defines agents as code: prompts in Markdown, tools in TypeScript, deployed on Cursor's hosted runtime or your own.

Anthropic has the Claude Agent SDK. OpenAI has AgentKit and the Codex SDK. LangChain has LangGraph. CrewAI has its own orchestration layer. Cursor's entry comes two months after xAI closed the acquisition that folded it into the Grok team, and it ships with subagent routing, hooks into the agent loop, and built in evals, so a team can measure whether a prompt change actually improved outcomes instead of guessing from vibes. It's TypeScript only right now, public beta, with tool call schemas unstable enough that Cursor tells developers to parse them defensively.

What's different about this entry is distribution. LangGraph and CrewAI are frameworks you choose to adopt from zero. Claude Agent SDK and AgentKit are vendor SDKs tied to one model family. Cursor's version sits inside an editor that's already open on a developer's screen all day, so the SDK doesn't need anyone to make a fresh decision to install it, just to notice a new menu item.

My read: the framework layer is consolidating around whoever already owns the workflow, not whoever has the cleanest abstraction underneath it. Owning the surface people already live in beats owning the better API.

https://t.co/3kug8qJm7n

[Michael Gabriel @mike_gabriel]

@edinsoncode Eh, I prefer advising our teams to make those kinds of tasks impossible to complete via restricting tools rather than via prompting.

For evals, I usually stick to golden sets based on real logs and edge cases that involve stopping early to ask for clarity.

[Michael Gabriel @mike_gabriel]

@larsencc On our PR descriptions*

[Michael Gabriel @mike_gabriel]

@larsencc 100%. I got my team to start requiring manual verification steps on our PRs and in increased our throughput like 2-3x since we didn't need to review most of the code to make sure shit worked.

Any tips on getting them to agree to ditch PRs altogether?

[Michael Gabriel @mike_gabriel]

@albertfranquesa Interesting choice.

At my own company, we've been moving a lot more stuff towards internal tools, primarily so we can use the logs to generate evals and gradually move towards open source / self-hosted models.

Wonder if they're doing something similar...

[Michael Gabriel @mike_gabriel]

@aroussi 100%. I'd also more explicitly measure tool retries and steps taken (even though they'd be part of latency) as a way to tell when something in the agent's environment or config is going wrong.

[Michael Gabriel @mike_gabriel]

Never understood people like this.

We're literally in the era where you can have micro apps built for you for $30/mo.

If the recomp is buggy as hell, fork it and get your own AI to fix the bugs you run into.

If it's free and open source, why get so triggered by it being AI?

[✧ ne0nyan 🪐 @ne0nyan]

vent art

[Michael Gabriel @mike_gabriel]

@quantian1 Around 3 and a half

[Michael Gabriel @mike_gabriel]

@thdxr Holy shit, assuming that's unique users and not devices, that means 1 in 5 coders use opencode now

[Michael Gabriel @mike_gabriel]

@mattpocockuk It's genuinely crazy how much stuff this applies to. Asking about principle of least privilege makes an AI suddenly go into security mode, asking about JOIN complexity makes it suddenly know how to write efficient queries, etc.

[Michael Gabriel @mike_gabriel]

@thdxr Finally, my dream of having subway surfers running on the side while vibe coding has come true

[Michael Gabriel @mike_gabriel]

I think my @grok @bot just paid for itself...

I used to pay $30/mo for a workout app that would automatically increase weights and reps for me based on hypertrophy research.

So, I got my bot to look into the research for me and build a micro app for my routine.

[Michael Gabriel @mike_gabriel]

@tmaxftw 1-2 years of pay to find a job + unemployment if you don't vs 2-3 months of pay.

It's also significantly easier to find a job when you currently have one and can say "7% layoffs when profits were up" as your reason for switching.

[Michael Gabriel @mike_gabriel]

IMHO, since they're offering to let up to 30% of their staff go, they shouldn't be allowed any H1-Bs. They obviously don't need the talent.

And why do we act like it's good for the immigrants? They almost never get green cards and are basically hostages who will get deported if the company decides they no longer need them. For that reason, they're less likely to ask for raises or complain about bad practices.

I understand we get people like @elonmusk 0.01% of the time, but this program only benefits people who don't want to stay and people who want to abuse those who want to stay.

[Michael Gabriel @mike_gabriel]

@tekbog Surely we could equip the drones with some sort of projectile package-delivery device to make delivery even smoother!

[Michael Gabriel @mike_gabriel]

@jerryjliu0 Honestly, no constructive feedback so far aside from saying great work! Main use case from my end is manufacturing/ heavy equipment manuals and it seems to do a great job with diagrams, multi-step instructions, and spec sheets.

[Michael Gabriel @mike_gabriel]

@jerryjliu0 From personal testing, it also works super well with assembly/disassembly diagrams.

Honestly, great work!