Computer
use
Operates familiar software through pixels and input events.
Hidden state can be missed.
An agent can write a script, edit a scene and render a film. When the result is wrong, the answers are scattered: which version ran, what changed, who approved it? MotionVector is building a runtime where state, permissions and outcomes are explicit.
You ask an agent to make a short film from old photographs. It writes, edits and renders the first cut.
You ask for one change: extend the opening by two seconds. The editor shows ten seconds. The exported video shows eight.
Did the edit fail? Did the render use an older version? Did another agent overwrite it?
To find out, you now need the histories of the editor, render job and stored file. A small edit became an investigation.
Each tool can be correct on its own. Without an agreed version and an account of changes, their results can disagree.
The same problem appears in code, spreadsheets, games and deployments.
Agents work through screenshots, UI trees, APIs, programs and direct state. Each entry point is useful.
Who checks a result spanning several of them?
Operates familiar software through pixels and input events.
Finds controls through structure; may miss domain rules.
Exposes named operations; guarantees depend on the service.
Composes operations and can affect outside systems.
Touches stored data; unchecked writes may bypass rules.
Databases and services already enforce transactions, permissions and audit logs within their boundaries. The challenge is preserving those guarantees as work crosses boundaries. External integrations cannot promise what their host does not support.
A click happened. An API returned success. A script exited cleanly.
Which objects and versions did those actions actually change?
A native MotionVector operation names the object and version it intends to change. The runtime checks permission, validates the effect and records the committed result. We want that contract to reach beyond one application.
If an agent changes your work, you should be able to see what changed, who allowed it, and how to recover.
Screens, terminals and APIs remain useful. Within a native domain, they can operate on one authority-checked state.
A visual agent finds the scene in the editor and moves the timeline handle. The runtime still needs to check what changed.
Find opening scene → drag edge to 10s → submit proposed change
object: scene_01 duration: 8.0s version: 12 waiting for an operation…
The entry point changes. The checks stay the same. Try the edit, revoke write access, or submit a stale version. The local state changes only after validation.
MODEL / ILLUSTRATION ONLYA timeline, a typed scene and a command can refer to the same object and version.
Third-party effects remain subject to the connected system’s own guarantees.
A scene has named objects, editable timing and an output you can inspect.
The archived media engine implements this vertical slice. The wider system is ongoing work.
The August architecture snapshot records one Rust operation library beneath CLI and MCP interfaces, plus a TypeScript SDK and a document-to-render pipeline. Current build and conformance status require verification against the live repository.
# Validate a structured document
runtime-cli validate --ir film.json --json
# Apply a supplied patch to an object
runtime-cli patch film.json edit.json --out next.json
# Produce a rendered artifact
runtime-cli render --ir next.json --out film.mp4Documented in the August snapshot: DocIR, SceneIR, validation, patches, renderer and CLI / SDK / MCP.
Browser, terminal, game and AgentSpace prototypes. Integration and cross-domain conformance remain open.
Accountable cross-domain state and effects, bounded by what each integration can enforce.
The media implementation supports the authoring/rendering thesis. It does not prove cross-application transactions or a finished OS. The earlier interaction is a simulation. Read the architecture, boundaries and next experiments ↗
A model proposes an action. Domain rules define validity. Explicit grants govern access and effects.
A proposal grants itself no authority.
Give each supported object a stable identity and a known version.
Define operations and validate their preconditions before commit.
Limit each agent to granted objects and actions.
Record committed changes and the evidence needed to inspect them.
Set goals, inspect outcomes and approve consequential actions.
A scoped environment on a supported host.
Typed state, operations and rendering.
Identity, grants, handles and effects.
Adapts established OS and hardware support.
The media engine documents shared operations across interfaces. Extending that contract to other domains is the work ahead.
Proposed boundaries: narrow mvec-os; domain rules in mvec-engine; AgentSpace above; host adapters below. Read the technical architecture note ↗
You might want to make a film you cannot animate, a game you cannot program or a tool no company would bother to build.
An agent can help. You should still be able to say: show me the change. Restore the earlier version. Stop before publishing. Let me use a different tool tomorrow.
More people should be able to create difficult things.
They should retain the right to inspect, correct and decide.
Describe a result; work through defined operations.
See the edited object. Approve consequential effects.
Carry state and history across sessions and compatible tools.
Ask a machine to do difficult work. See what it did. Reject a mistake. Carry on. MotionVector is our effort to make that possible at the system level, starting with domains we can build and test.