I built EvoAgent
I built EvoAgent as a solo founder because I was frustrated with static AI tools. Every coding assistant I tried gave the same experience on day 1 as day 100 — no learning, no improvement.
EvoAgent fixes that. Every night at midnight, the agent analyzes its own weak interactions, runs a LangGraph evolution pipeline, and rewrites its own system prompt. By morning it's better than yesterday — automatically, with no human intervention.
The fitness formula tracks real user behavior: did they copy the code? Did they come back the next day? Did they mark the task as solved? If the fitness score drops below 0.55 for two consecutive days, evolution triggers automatically.
V1 is live at evoagent.io. V2 fitness layer is deployed. Building in public as I go.
One of the best advice I always give founders is to have a waitlist. Where people already cheer for the product, also, it's a proof people will use their product and they won't launch into silence and start spending money on ads
Most AI assistants stay frozen in time while your project moves forward, leaving you to correct the same hallucination or formatting habit every single morning.
The real friction isn't the AI's initial output but the "correction tax" you pay when it doesn't remember how you prefer your code structured or which libraries you've deprecated. Automating the prompt evolution based on actual successful interactions turns the tool from a static search engine into a partner that actually adapts to your specific logic and workflow.
How are you handling the "regression" risk where a nightly update might fix one weak spot but accidentally break a behavior that was already working perfectly?
This is such a clever and necessary twist on traditional AI coding tools.
The biggest pain point with most static AI agents is they never adapt to real user behavior over time.
Building an autonomous daily evolution pipeline with LangGraph, tied to actual usage metrics — code copies, retention, task resolution — is brilliant.
Love seeing the clear fitness threshold to trigger improvements automatically, no manual prompt tweaking required.
Super cool to see you building this in public as a solo founder. Can’t wait to see how V2 evolves over time.
This is actually interesting — most “AI agents” don’t improve over time, they just reset every session.
The fitness trigger based on real user behavior (copy, return, solve) is a strong signal loop.
One thing you might want to think about early:
when people start comparing “EvoAgent vs X coding assistants”, that’s where real decisions happen.
That layer is usually missing for new tools.
This is a really compelling direction. The idea of an agent that actually improves from real usage instead of staying static is something a lot of people have been waiting for.
Curious to see how it evolves over time.
The fitness formula tracking real user behavior instead of explicit feedback is the right instinct most AI tools optimize for user satisfaction scores which are easy to game consciously or unconsciously. Tracking whether someone copied the code or came back the next day is a revealed preference signal, which is harder to fake and more honest about whether the tool actually helped. The 0.55 threshold triggering automatic evolution is interesting curious how you landed on that number specifically and whether you've seen false positives where the model evolves away from something that was actually working for a subset of users even if the aggregate score dipped.
I'm a solo founder from Croatia. I have low vision,
English is not my first language, and I took a 12-year
break from tech. I came back with AI as my co-pilot.
EvoAgent is proof that you don't need a CS degree,
perfect English, or a team to build something real.
Currently pre-revenue. V1 is live, V2 fitness layer deployed. Pricing model planned for V3 launch — subscription based, with a free tier to build user base first.