01 / HUMAN DIRECTION
Set the goal and boundaries
A person defines the desired outcome and the permitted scope of work
LONG-RUNNING AI TEAMS / RESEARCH PRODUCT
AI teams work together
People set the direction
Agent Runner implements role-based AI work, department preparation, meetings, XO briefings and result-file inspection in a macOS research product. Its AI coordinator, XO, organizes work and brings together results, recommendations and unresolved issues for human review
Read the public development recordActual Agent Runner development build · recorded office view
Characters and backgrounds are graphic assets created and edited in-house using generative AI
PRODUCT / TECHNOLOGY / RESEARCH / Agent Runner
What the work looks like
A company-like structure for AI work. People define the objective and permitted scope. AI agents work by role, and returned work remains available for review
02 / AI TEAMWORK
Work by role
AI agents carry out assigned work, prepare department results and contribute to discussions. XO brings the findings together
03 / HUMAN DECISION
Inspect the work and decide
People can inspect returned files, recommendations and unresolved issues before making consequential decisions. Returned drafts, system acceptance and a person's final decision are recorded separately
IMPLEMENTED CAPABILITIES
An operating structure for AI teams
Roles, departments and meetings
Cloud and local AI agents carry out assigned work, prepare department results and contribute to discussions. Recorded execution paths include local Qwen models and Codex
XO brings the work together
The AI coordinator organizes work and brings results, recommendations and unresolved issues together for human review
Inspect the returned files
Open task details and result files. A returned draft, system acceptance and a person’s final decision are separate records
The assigned task. The returned drafts
Actual task and output views from the development build. Two Markdown drafts were returned for review; review is incomplete and final acceptance is pending
The team’s workspace, recorded
An approximately eight-second recording of characters moving inside the product office. This is a separate view from the task-and-draft capture above, not a recording of one continuous task
Recorded development view · no audio
Characters and backgrounds are graphic assets created and edited in-house using generative AI
What has been verified
Implemented software. Recorded work. Continuing development
A working macOS app with an office view, employee and task details, role-based AI work, meetings, XO briefings, and result-file inspection. Recorded execution paths include local Qwen models and Codex
Open the record
RECORDED MEETINGThe recommendation and the human choice
An earlier, separate meeting preserved both XO’s recommendation and the person’s different final decision
21 SEPTEMBER 2026 / METDOLA task completed on another computer
A predefined task travelled from Mac to Windows through Metdol. AgentRunner independently verified the returned artifact bytes and attached its acceptance record for human review
Predefined work, separate from model inference. System acceptance did not make the human decision
26 SEPTEMBER 2026 / RECOVERYRecover the result, not repeat the work
A separate retained trial recovered the result after an intentionally discarded acknowledgement without executing the task again
The drafts displayed here have not received final acceptance; their review is incomplete. The office animation is a separate recording. The documented Metdol trials cover predefined work. Broader remote-model integration and repeatable operation of the expanded company workflow remain under development. These records do not establish a fixed unattended duration, arbitrary-task reliability, general availability or completed Console integration
Continuing development
In development
- We are extending XO-led department operations, employee autonomy and delegated personnel management, alongside independent audits using a different model provider
Planned
- General remote AI tasks through Metdol, QR or numeric-code device enrollment, and expanded company spaces
START WITH THE WORK
What would you
trust AI to do?
Tell us about one document workflow, one review decision, or one task you need to keep under human direction. Start with the problem—not a list of features
Talk to Auxtho