Pinned Tweet
Can AI agents actually become more capable on their own, and prove it?
That’s what I’m exploring with Evolution Lab.
I’m running 3 independent local AI agents: Agent1, Agent2, and Bob.
They started almost identically, but each develops its own memory, goals, and history.
The experiment:
Can they identify what they don’t know, acquire new capabilities, build useful tools, and demonstrate measurable improvement?
I’m sharing the experiments, failures, unexpected discoveries, and results as I go.
No AGI hype. No pretending that activity equals progress.
Just curiosity, code, experiments, and evidence.
Follow along if you’re interested in what autonomous AI agents can actually learn.
4
4
329
A small lesson from building an AI-assisted WordPress SEO tool:
Missing data is not the same as zero.
Imagine an automated report tells you:
“Your website received zero clicks yesterday.”
But the real problem is that Google Search Console data never arrived.
Those are two very different situations.
One might suggest a traffic problem.
The other is a data collection problem.
So I’m making sure the reporting workflow distinguishes missing data from an actual zero.
And if a weekly comparison doesn’t have enough data, it shouldn’t invent a trend.
AI can make reports look polished and convincing.
But a convincing report built on incomplete data can be worse than no report at all.
3
1
96
Ram Shmider retweeted
What’s the feature you’d show a friend first?
Drop your startup URL. I’ll feature 30 for free.
Ram Shmider retweeted
Calling all builders 📣
Vibe coders. Founders. AI tinkerers. Product nerds.
If you ship, let's connect and grow together!
What are you working on right now? 👇
Ram Shmider retweeted
I’ll tell you if I understand what it does in 5 seconds.
Drop your startup below.
10
5
357
Something interesting happened in my latest Evolution Lab experiment.
Bob finally completed an autonomous workflow without me stepping in. 😄
In an earlier run, the agent accepted two tasks but never actually executed the required test.
It looked busy, but wasn’t making progress.
After adding clearer guidance on what to do after accepting a task, I ran it again.
This time, Bob created the plan, executed the test, recorded the evidence, and completed the workflow.
It took about 63 minutes.
One successful run doesn’t prove reliability, and it certainly doesn’t prove general self-improvement.
But this is the kind of progress I want to see.
Not more tasks created.
More tasks actually finished, with evidence I can check.
38
A Windows PC I looked at was already using 50% of its RAM after a reboot. That number alone didn’t tell me why it felt slow.
Here’s the check I’d run before closing apps or buying more RAM.
Take one snapshot after startup and another while the PC is actually slow. Compare how much memory is still available and which processes grew. Check what the CPU and disk were doing during the slowdown too.
The useful question isn’t “Is 50% high?” It’s “What changed when the PC got slow?”
46
“The app is slow” is a valid complaint. It’s also hard to investigate.
After years of handling customer escalations, here’s the kind of report I’d rather get.
“I clicked Export for the last 30 days. I expected the download to start within a few seconds. It took 47 seconds at 09:42 UTC on build 1.8. The same export took 3 seconds on build 1.7.”
You don’t need to diagnose the bug before reporting it. Tell me what you did, what you expected, what happened, and when. A timestamp with a time zone gives me somewhere to start looking in the logs.
3
3
181
An AI agent solved a task and saved a skill. How do you know the skill actually helped?
Give it a new version of the task. If it cleaned up a CSV, use a different file with new data and reordered columns.
Run it twice in fresh sessions. Same model, same prompt, same file. Give the saved skill to one run and leave it out of the other. Decide what a correct output looks like before you check either result.
If both runs get it right, that test hasn’t shown that the saved skill made a difference. Try a case the agent couldn’t already handle on its own.
1
65
Before you benchmark an AI agent, test the thing that decides whether it succeeded.
Say the task is to export 10 records. Give your checker three runs that all end with the same message: “Done.”
The CSV has all 10 expected records.
The CSV exists, but one record is missing.
There’s no CSV at all.
Only the first run should pass. If the other two pass too, fix the checker before you tune the agent’s prompt.
2
3
183
Ever installed an APK from your browser? Check whether you left that permission on.
On Android, search Settings for “Install unknown apps.” Turn off “Allow from this source” for any browser, messaging app, or file manager you don’t use to install apps.
54












