Halo Architecture and Runtime
1. The workspace shape​
The README reveals a system built from several distinct ideas:
| Area | Why it matters |
|---|---|
| workstation product layer | the main user-facing shell |
| pluggable agent engines | connect Claude Code, Codex, and related backends |
| digital human subsystem | runs scheduled autonomous work |
| browser actions | reusable deterministic scripts for web tasks |
| stores and distribution | make workflows reusable and organization-friendly |
That is a large product footprint compared with a normal coding CLI.
2. The runtime mental model​
Halo generally works like this:
- the user or scheduler defines a task,
- an agent engine reasons about the task,
- browser actions or tools perform stable operations,
- outputs return into the workstation,
- remote-control surfaces can monitor or steer the process.
3. Why browser actions matter​
Halo explicitly argues that AI should decide what to do while prewritten browser actions handle how to do it. That design matters because it reduces the brittleness of pure free-form browser agents.
4. Digital humans are a first-class subsystem​
The project treats autonomous scheduled workers as a real product category, not as a thin cron wrapper. That is one of the clearest differences between Halo and ordinary desktop AI apps.
5. What to read first in code​
Start with:
- the README product overview,
- browser-action and digital-human docs,
- desktop and runtime structure,
- integration and store surfaces.