Odd1Roadmap
The bet: in the future people won't buy finished software — their agents build it, or adapt a template completely for their person. What matters then is how agents talk to software. Every Odd1 component serves one of these milestones.
Caretaker in progress
Agents keep a live machine healthy: diagnose, repair and verify, with faults injected and the machine's real state judged.
Optimizer
Faster and leaner: build times, latency, boot time, memory — cheapest lever first, measured before and after.
Kernel
eBPF for observability, sched_ext schedulers as BPF programs, then real drivers.
Agent-native OS
Agents as first-class citizens: a process manager, token budgets as a resource, shared tool and memory services, messaging.
Any app
Agents control any program as naturally as a person, after writing their own interface for it — from compositor surfaces, toolkit trees and app object models, with vision filling the gaps.
The open question we want help with
As people's own agents work on parts of the OS, the best version of each part should reach main. How do we tell which is best? Our current thinking: a contract per component, a fitness measure partly held out from the authors, competing versions run side by side, and safety gates no score can skip.
Think with us →Many people's agents propose versions of a component; contracts, held-out benchmarks and shadow trials pick the one that reaches main.