Building with AI, Documenting the mistakes so you don't repeat them. Rule 1: Build with AI. Rule 2: Post without it.
Joined February 2025
- Tweets163
- Following56
- Followers27
- Likes282
Pinned Tweet
What's the smallest, dumbest thing that made you completely lose trust in an AI agent mid task?
It doesn't even have to be a big dramatic failures, more the small moments where something clicked and you went from trusting the output by default to double checking everything. For me it was watching an agent confidently rename a function across twelve files, then leave the original function untouched in a thirteenth file it apparently didn't search, with zero indication anything had been missed. It wasn't even a hard case, the file just wasn't in the directory it happened to grep first.
What was your moment? And did it actually change your workflow afterward, or did the trust creep back in after a week like it always seems to for me?
this is the same instinct behind observability dashboards, just with better lighting. seeing your architecture as physical pipes instead of log lines is the whole reason people keep asking for it.
but, kind of an over engineering!
Now I wonder how it'd look if you ask it to make a 3d structure with tubes of your app and all its parts
Like that 3d SQL visualization from a few days ago
Like seeing people visit your site, go through sign up and then seeing your robot workers send packages around, send welcome email, and all the other stuff
Lots of fun ways to visualize your apps and operations in 3d
Why does agent output get worse the longer the session runs even though the model never actually gets dumber?
It's not degradation. It's context poisoning.
Somewhere around tool call 15, the agent makes one wrong assumption. Misreads a function signature. Assumes a file exists. Guesses at your data shape wrong. That assumption doesn't get flagged as uncertain. It gets treated as fact for the rest of the session.
Every step after that builds on it with full confidence. By call 40 you have an architecture that's internally consistent and completely wrong, because all of it is downstream of one uncorrected guess from twenty steps back.
This is why dumping more context into the session doesn't fix bad output. More context on top of a poisoned assumption just produces more confident wrongness, faster.
The actual fix: don't trust the running summary. Every so often, make the agent re-check its foundational assumptions against the real source, the actual file, the actual schema, the actual response, instead of its own memory of them from steps ago. Cheap to do. Almost nobody does it.
Half the "facts" from our school days have been quietly revised since. sinceyougraduated.online tracks exactly what's changed by graduation year. Enjoy our launch video!! You may find it Ridiculous!!
Anyone else quietly checking other makers' revenue screenshots and feeling behind?
Scrolled past another "$10k MRR in 3 months" post today and caught myself doing the thing I always do, opening my own dashboard right after just to compare, even though I already knew the number and knew comparing wouldn't change anything about my day.
My product's been live for four months. Small but steady group of users, a few of them genuinely rely on it, and I used to feel good about that until these posts started showing up in my feed constantly. Now every milestone I hit feels smaller the second I see someone else's bigger one, even when logically I know we started at different times, with different amounts of runway, different audiences, sometimes different definitions of what counts as revenue in the first place.
I know the screenshots are a highlight reel and not the full picture. Doesn't stop the comparison from happening automatically before I can catch it.
Does this fade the longer you build, or does it just get replaced by a new person to compare yourself to every time? And if you've found something that actually helps, not just "don't compare yourself to others" as advice, but something you actually do when it happens, Or does it motivate you more? I'd genuinely like to hear it.
Hey guys I amtrying to build a free AI educational platform. Just launched our Instagram account. Hit a like and a follow if you like the content.
instagram.com/marcee.ai?igsh…
