Last updated March 8, 2026
The past few months crystallized something: I’m an agent infrastructure builder now. Not in the “slap an LLM on a chatbot” sense—I mean the actual runtime systems that make agents work reliably in production.
At Workus AI, Marco has matured into a solid Agent Runtime. The Domain Pack architecture, the Skills system, multi-agent management—it all works. We’ve moved past the “will this even work?” phase into “how do we scale this?” territory. That’s a good place to be.
But the more I built Marco, the more ideas kept nagging at me—ideas too experimental for a production system but too important to ignore. So I started Alan, an open-source agent runtime that I’m building as the foundation for all my future agents. The core question Alan explores: what if we model agents as Turing machines, where human involvement shifts from operating the agent to defining its goals and owning its outcomes? I wrote about this in my Human-in-the-End post.
Marco is the pragmatic one, shaped by production constraints. Alan is the long-term bet—where I can pursue the architecture I actually believe in without compromise. They feed each other, but Alan is where I’m investing my future.
I’ve come to a conclusion that’s been brewing for a while: don’t build tools for humans—build tools for agents.
Kira was supposed to be an AI-native second brain—capturing context, connecting ideas. But in an agent-first world, that’s just a capability an agent should have natively. Kira’s entire value proposition dissolves into a feature of something like Alan. The project was wrong from the start, not in ambition but in framing.
Sealbox is paused for a similar reason, though the logic is slightly different. Secret management will evolve beyond API keys and vaults. I suspect we’ll see agent authentication mechanisms closer to biometric identity than to shared secrets. Building a “better HashiCorp Vault” makes less sense when agents will manage their own credentials.
The insight isn’t that human-facing tools are worthless—it’s that personal project time should be invested in agent infrastructure, where the leverage is highest. openrouter-rs is still a developer SDK, and that’s fine. But Kira and Sealbox were trying to be end-user products in a world that’s shifting underneath them.
This one felt like a milestone. openrouter-rs reached 100% coverage of the official OpenRouter API—all 36 endpoints, fully typed, fully tested.
The 0.6.0 release cleaned up all the 0.5.x compatibility baggage: deprecated aliases removed, legacy completions moved behind a feature gate, domain-oriented client API solidified. Six clean domain clients—chat(), responses(), messages(), models(), management(), legacy()—each mapping directly to the official API surface.
I also shipped openrouter-cli, a full command-line tool for interacting with OpenRouter: model discovery, API key management, guardrails, usage tracking, credits. Table and JSON output, multi-profile config support, the works.
Going from “a Rust SDK with some endpoints” to “the definitive Rust interface for OpenRouter” felt good.
What started as “learning Godot for fun” has turned into a fully playable prototype. 16 characters, complete battle system, training mode, story mode, AI opponents, and an AI-powered art generation pipeline.
The game is a satirical Silicon Valley brawler—Elon Mvsk vs Mark Zuck vs Sam Altmyn, each with signature moves based on their companies and products. The combat system has real depth: combo scaling, hitstun, guard counters, throw techs, knockdown recovery. It’s not Street Fighter, but it’s not a joke either.
Recent work has focused on the onboarding flow and menu system polish. The pixel art UI came together nicely with a cohesive skin pack. Dual language support (EN/ZH) is already in.
Still a side project, still purely for fun—but I’m proud of how far it’s come from “basic scene setup” three months ago.
The band idea from last month? Dead. Nobody had the energy. Fair enough.
More Marco at work. More Alan experiments in the open. Maybe actually publish openrouter-rs 0.6.0 to crates.io. And keep punching Silicon Valley founders in my spare time.
The work continues.