Intent
Define the goal, boundaries, and expected artifact.
VMLogic Harness / macOS
In development · No public release dateA native environment where agents remember, act, and leave evidence.
VMLogic Harness is the planned macOS product surface for persistent agent work. It is designed to connect durable project state with apps, browser sessions, files, and tools—then make the resulting actions and updates inspectable.
PlatformmacOS
ArchitectureModel-flexible
StateIn development
TaskBuild an evidence-backed market brief
01
Relevant project state recovered
02Claims mapped to sources and constraints
03Browser · Files · Analysis · Document
Artifactbrief.pdf
Verified artifact · trace retained
System loop / 01
The work continues beyond one model response.
A harness coordinates the surrounding system: what should be remembered, which tools may act, how progress survives, and what evidence supports the final result.
Context
Recover only the project state relevant to this decision.
Plan
Turn relations and constraints into inspectable work.
Execute
Work through authorized apps, browser sessions, files, and tools.
Verify
Check the artifact, sources, actions, and unresolved questions.
Persist
Store evidence-backed state so the work can continue later.
One workflow / Research brief
From a request to a continuable workspace.
The example below is a product-direction scenario, not a claim of current public availability.
Request
“Compare three approaches to persistent agent memory. Produce a concise brief, preserve the evidence, and leave open questions for the next session.”
Recover the research state
Load the relevant project, prior decisions, source register, and outstanding questions.
Project stateCollect and structure evidence
Use the browser and local files, mapping each claim to its source and timestamp.
Evidence graphCompose and check the brief
Draft the comparison, test constraints, and surface conflicts rather than smoothing them away.
Reviewable draftDeliver and preserve continuity
Write the artifact, retain the execution trace, and carry forward only supported updates.
Brief + next stateProduct architecture / Design targets
Built around the seams where agent work usually breaks.
Persistent project context
Projects, people, decisions, evidence, and open questions can remain available across sessions without replaying an entire transcript.
A real macOS work surface
The harness is intended to coordinate authorized work across local files, apps, browser sessions, and external tools.
Long-running and recurring work
Checkpoints, resumable progress, and scheduled routines are treated as system concerns rather than prompt conventions.
Skills and deliverables
Repeatable methods can be packaged as skills, while results are delivered as real documents, data, code, and other artifacts.
Inspectable by design / 04
Persistence should not become invisible automation.
The product direction is to keep cognitive state and execution reviewable: where context came from, what acted, what changed, and what remains uncertain.
Memory provenance
Recovered state should retain its source and scope.
Execution trace
Tool actions and produced artifacts should remain inspectable.
Human checkpoints
Consequential actions should expose clear review and control points.
Model boundary
Durable state should remain separable from the selected foundation model.
Supported models, exact permissions, data handling, availability, and pricing will be published in formal product documentation before release.
VMLogic Harness / Product direction
The native product surface for persistent AI.
In development for macOS. No public release date has been announced.