Magine evolved from a simple GitHub profile generator into an AI agentic platform featuring sight-driven agents and a distributed GPU mesh, with growth metrics and technical innovations in agent orchestration.
Yes, you saw it right: a GitHub profile generator that accidentally traded stocks, started trading then became an agent orchestrator platform. Now we're trying to make coding agents measure instead of guess. 7 months ago I build a small tool that generated a GitHub profile card (only for me). That's it. A weekend thing. Somewhere in a private branch of that repo, I added an agent orchestrator feat to it. Not for anyone else. I wanted something that would watch ETFs & trade itself while I slept and act on rules I'd written down. It broke. I patched it. It broke again. After a few more rounds it did better than I expected and generated a huge profit that month, well enough that I stopped treating it as a toy. What I didn't have was a way to see what it was doing. Every "agent builder" I looked at was the same thing: a chat box, a sidebar, hundreds of MCPs integrators, connect dozens of APIs, a copy of a copy. I didn't want to build another one of those, and I didn't want to use one either. So I built the opposite. A single page. A retro terminal. Rather than using playwright blindly let settle it down to use vision based approach which we names as Sight Driven Agents (SDAs). You type, agents run, and you watch them work in a live sandbox like a kid at a desk. No dashboards to click through, No MCPs, No connectors (OpenClaw hater), IT ACTS AS YOU. It still looks like that today, and it's going to stay that way. That's how Magine 😺 started. Here's where it is now, and the part I actually want to talk about: why we scrapped our coding agent. Where we are $1.2k MRR (low, and partly on purpose: we hand out a lot of freebies to our beloved early users 😽), growing ~19% month over month Zero marketing spend. Zero ads. Every user came from a direct invitation or a beta invite 5,000+ autonomous agent hours run on the platform 7M+ browser frames captured from agents working inside real browsers 1,000+ concurrent browser containers at peak Grew from team of duo to 4. I'm still the only technical person Left my CTO job and took on a day job just to keep the lights on. That means I have almost no time left to balance core development with the growth-critical meetings we desperately need right now. Meanwhile, my inbox has dozens of emails from VCs interested in our MRR-sorry, our **product, **while we take casual intro chats ~40% of what we burn goes to GPUs We give every new user 1M tokens to test with. That might not sound like much if you're already familiar with other platform, but because we aggressively cache context and compress inputs and outputs, those tokens go noticeably further than the raw number suggests. Two things I'm quietly proud of. First, Magine features triple-redundant execution: if Vision fails (Blame GPU availability), it defaults to CDP; if CDP fails, it leverages classic devtools and VLMs to combine frames, mirroring the approaches of Anthropic and Manus. Additionally, our roadmap ensures Magine will operate entirely offline using edge LLMs. Second, everything heavy runs on a mesh of gpu worker machines we enroll from the platform itself: LLM inference, vision inference for the browser agents, model training, and the sandboxes CatCode uses to measure code. Work is routed to whichever node has the GPU, RAM, and headroom for it, inference fails over between nodes, and a sandbox gets its own isolated slot instead of fighting the model for the same card. The 1,000+ concurrent browser containers above ran on that mesh. We dropped the traditional coding agents this month We had one. Open a project, spin up a sandbox, tell the agent what to build. It worked. It was also completely uninteresting, because ten other products do exactly that, and none of them answer the question that actually matters when an agent edits your code: Did it get better, or does it just look different? An agent that says "I optimized your function" is giving you an opinion. Nobody should ship an opinion. So we rebuilt the thing, and we call it CatCode. Surface-level, this is how it works: You give it one sentence. "Make the matcher in src/match.py faster without changing its output." No config, no markers, no setup. It picks the code worth changing and asks you to confirm before it does anything. You can say yes, no, or redirect it in your own words. It follows your words over its own plan. It writes its own evaluator, then tests the evaluator by feeding it a deliberately broken version of your code. If the judge can't tell broken from correct, the run stops. It will not optimize a number that means nothing. It generates competing ideas from several different angles, scores them before writing any code, turns the best into candidates, races them through a cheap screen, and fully measures only the survivors. Inside your own sandbox. It measures your original several times to learn the noise, re-measures the winner, and only calls it a win if it clears the noise on two independent statistical tests. It writes a DISCOVERY .md into your repo: baseline runs, confirmation runs, the exact diff, how running time scales with input size, and a threats-to-validity section. Every number in that file was measured, not estimated. It also never provisions or resizes infrastructure by itself. If your sandbox is too small, it tells you the exact size the job needs and the one command to get it. Your machine stays yours. This is the AlphaEvolve family of ideas (propose, measure, keep what wins) pointed at your code, with a judge that gets audited before it's trusted and an agent that asks before it acts. That's the direction. We are not claiming we've beaten anyone's numbers. The research side Besides the orchestration part, we've been working on a JEPA-inspired architecture we call RD-S4 / EFR, an attempt at an attention-free, linear-complexity model that adapts its internal pathways while it runs. It's early. We're preparing a paper and will share it when the results hold up, not before. What's next Video generation using vision-based workflows that track objects and subjects inside the platform Open-sourcing Magine once CatCode leaves experimental status. It's still experimental. We're still drawing conclusions about how well it performs, and we'd rather show you that honestly than early A marketing workflow, because right now there isn't one, and the growth you see above is entirely word of mouth Where I genuinely want readers advice Two decisions I keep going back and forth on, and I'd rather hear it from people who've shipped than guess: Video Generation. We tried building a movie-like workflow once already: camera angles, shot direction, classifying the objects and subjects in frame so an agent could track them across a scene. It failed. Not "needs polish" failed, properly failed. We're about to attempt it again with the vision approach that runs our browser agents, and I honestly don't know if that's persistence or stubbornness. If you've built anything in video generation or object tracking, tell me what we're walking into. Scope. Every platform around us is compacting everything into one product, and this is us doing the same. The alternative is to drop it and put the whole team behind the two things that are actually ours: the model research and vision-based browser automation. Is "one platform for all of it" the right call for a 4-person team with one engineer, or is it how small ideas die? I'll take the honest answer over the polite one. And if you've read this far: thank you for reading this journey 😊. I want every one of you to be part of it, and I'll keep posting exactly what we deliver, including the parts that don't work.
A viral GitHub repository called "The Agency" offers 155,000+ stars worth of pre-built, personality-driven AI agent personas (engineers, designers, marketers, etc.) ready to install into Claude Code, Cursor, Codex, Gemini CLI and more, with a companion desktop app for one-click setup.
MagesticAI is an open-source, browser-based multi-agent AI coding platform that uses Planner, Coder, and QA Reviewer agents in coordinated sessions with isolated git worktrees and supports multiple LLMs including OpenAI-compatible endpoints.
The author built an open-source, MIT-licensed platform for creating and managing AI agents, featuring provider-agnostic support, MCP integration, memory, skills, scheduled triggers, and Kanban boards, deployable via Docker Compose.
Agenvoy is a Go-based personal AI assistant that unifies Claude, GPT, and Gemini into a single productivity layer, allowing users to chat across multiple channels and automatically generate tools via natural language commands.