Last updated February 13, 2026
I kept telling myself I’d write a 2025 year-end review. It never happened. January came and went, and now it’s mid-February. The review is officially overdue, probably permanently.
The truth is, the past two months have been consumed by Marco. Not incremental improvements—a fundamental rethink of what we’re building.
The biggest shift: moving away from building yet another vertical B2B agent (sourcing, sales, whatever). I became convinced that the ceiling for point solutions is too low, too fragmented.
Instead, I’m now building Marco as a general-purpose Agent Runtime—something closer to Codex or Pi in philosophy. The core insight is separating the execution engine from business logic:
Domain Pack Abstraction: The same runtime core (session loop, tool loop, event streaming) can drive completely different behaviors through pluggable Domain Packs. Sourcing, Sales, Ops—they’re just different packs on the same engine.
Skills System: Business capabilities as hot-swappable skills, not hardcoded logic. The agent loads what it needs, when it needs it.
Multi-Agent Management: AgentManager/AgentInstance patterns supporting concurrent specialized agents, each with isolated state and memory.
The technical foundation is solid—Rust’s type system actually helps model these boundaries cleanly. But more importantly, it feels like the right layer to be building at. Less chasing feature parity with competitors, more creating infrastructure that could outlast the current wave of AI applications.
Life has been stable. Same routine, same rhythm.
One new thing: some colleagues and I are planning to start a band called MAJE. Nothing serious, just something fun to balance the intensity of startup life. We’ll see if it actually happens.
The work continues.