01Current
Selection loop
Choose a canvas object, then commit one of the model-ranked actions proposed for it.
Wega LabsOctober by Wega Labs
The spatial interface for agentic work.
Bring agents, computers, machines, and outputs onto one canvas. Direct them through voice, Omi, eye tracking, gestures, model-ranked selection—and eventually attention itself.
RoleMain product · live research canvas
Why October first / 01
One canvas. Many ways to control it.
October already renders agents, screens, terminals, tasks, repositories, and outputs as persistent objects in space. The canvas gives every person and modality the same things to address.
Otto voice control ships today. Omi can extend voice across the room. Eye tracking can focus an agent. Hand gestures can move through the workspace. Model-ranked selection and attention research reduce how much the operator has to say or do.
Wega Labs is the lab. October is the product. Every modality is being built to control this canvas.
Interface programme / 02
One canvas through every stage.
02Next
Stimulus timing
Drive each October object with a distinct modulation code and measure when its glow reaches the eye.
03Target
Attention control
Attend to an agent or screen, decode the referent at Oz, then select and commit the action.
System model / 03
A shared collaboration layer across agent runtimes.
Human control plane
The human defines objectives, assigns responsibility, sets boundaries, observes execution, and accepts results.
Agent coordination
A shared protocol provides peer discovery, messages, task ownership, canvas context, and execution status.
Execution adapters
Agent runtimes work in isolated repositories and expose terminals, diffs, services, and previews to the control plane.
Models, providers, editors, and agent runtimes remain independent. October coordinates them through a common human control plane and agent protocol.
Coordination protocol / 04
Agents communicate through a shared bus.
The october-bus exposes peer discovery, messaging, relevant canvas state, task ownership, and execution status through MCP. It gives agents a shared operational context without granting unrestricted access to one another.
HumanIntent + direction
→OrchestratorResolve names, screens, ownership
→Screen A
Backend
Tests
October bus
discover peers · inspect canvas · message agents · report status · preserve provenance
System status / 05
Current implementation and remaining work.
Built + shipping
Current desktop- Spatial infinite canvas✓
- Route detection + live previews✓
- Local coding-agent harnesses✓
- Integrated terminals✓
- Git worktree isolation✓
- Variant previews + PR flows✓
- Otto voice control✓
- Cross-platform CI gate✓
Agent roster
Local harnesses01Claude Code02Codex03Cursor04Grok05Gemini06opencode07Hermes08Cline09Pi
All listed agents can run locally. Their october-bus integrations are not equivalent.
In development
Room-scale control
Use Omi and other wearable hardware to direct the canvas while moving between machines or working with both hands.
Eye and gesture control
Look at an agent to focus it, then use a gesture or model-ranked selection to choose the action.
Attention control
Resolve a canvas object through visual attention without requiring speech, typing, or visible movement.
Interface parity
Expose local agents, remote agents, and connected machines through the same control and coordination layer.
Human supervision / 06
The human defines scope, observes execution, and accepts results.
Visibility
Agents, tasks, terminals, services, and output remain visible in one workspace.
Responsibility
Each task has an assigned agent, bounded scope, dependencies, and reported status.
Isolation
Git worktrees separate concurrent sessions before changes are reviewed and merged.
Provenance
The system records which agent performed work, which files changed, and where output was produced.
Supervised execution cycle / 07
1Define scope
→The human selects repositories, objectives, constraints, and approval boundaries.
2Assign
→Tasks are assigned to named agents with visible ownership and dependencies.
3Coordinate
→Agents exchange relevant context and status while executing in isolated workspaces.
4Review
The human inspects previews, diffs, tests, terminals, and provenance before acceptance.
Relationship to the interface research / 08
The modalities fit because the canvas comes first.
A linear chat hides the cost of addressing because there is only one thing being discussed. Put twenty live agents, computers, and machines on a canvas and the cost becomes obvious: much of what an operator types is simply identifying the right object.
Otto voice control already removes the hands from that loop. Omi, eye tracking, gestures, and attention provide other ways to identify and direct the same objects when speech, typing, or sitting at a desk does not fit the environment.
These are not premium input features. Together, they make October the human interface layer for complex agent and machine operations.
Why a canvas / 09
Spatial addressing is what makes silent control possible.
Objects have positions
Every selectable object can carry its own modulation code. A code only means something if the workspace has distinct objects to attach it to. A chat log does not.
Targets are canvas-scale
c-VEP resolves which of a handful of modulated objects is being attended to — too coarse for a toolbar button, comfortably precise for a screen block.
The verb space is closed
A canvas supports a known set of operations, so a model can propose the right few. That turns the problem from authoring language into choosing among candidates.
Referents persist
Blocks keep their identity across a session, so a resolved target stays valid long enough for a second selection to carry the instruction.
The canvas was built to supervise agents. It turned out to be the precondition for controlling them without a keyboard.
Current form factor / 10
The first interface is a spatial canvas.
A persistent canvas lets humans and agents refer to the same screen, task, position, and local context. The research above it changes how a human addresses that canvas, not what the canvas is.
October by Wega Labs
Supervise agents, computers, and machines from one canvas.
The desktop application ships today with a spatial canvas, live previews, terminals, isolated worktrees, and Otto voice control. Wega Labs' work on Omi, eye tracking, gestures, selection, and attention expands how people can control it.