AIOS 108ALPHA

Part of the ALPHA suite

ALPHA Agents · Delegate

Delegation with the leash attached.

Hand well-defined tasks to specialist agents that act only within role-scoped permissions. Every action is attributable, consequential steps wait for your confirmation, and any agent can be paused or revoked.

Request access Enter ALPHA Studio
Four narrow lanes converging on one approval gate

Roster pattern

Four archetypes, each with a lane.

Agents are assigned by role, not by personality. Each archetype carries a narrow grant, a stated approval line, and a readable record.

  • Role · Build

    Builder

    Drafts and edits inside a Studio workspace: new pages, revised flows, corrected content. Proposes changes for your review; applies nothing consequential without approval.

    Typical grants: workspace files under review, preview runs. Asks approval before: applying edits outside the brief, publishing, sharing.

  • Role · Review

    Reviewer

    Checks work prepared by others — people, builders, or runs: consistency, scope fit, declared permissions. Reports findings with references to the exact change under review.

    Typical grants: read access to the change set and its record. Asks approval before: altering the work itself, widening the review scope.

  • Role · Investigate

    Researcher

    Gathers and organizes material for a DeepResearcher brief: sources with provenance, comparison of viewpoints, uncertainty marked. Cites origins; never invents them.

    Typical grants: stated sources and workspace notes. Asks approval before: contacting anyone, publishing findings, expanding the question.

  • Role · Operate

    Operator

    Carries out the declared steps of an automation run: filing, staging drafts, preparing notifications. Holds at every approval gate for a person to decide.

    Typical grants: the run's declared tools and fields. Asks approval before: sending, deleting, sharing, spending, or changing the run itself.

How delegation works

Define the assignment, keep the decision.

You state the task, the limits, and the stop conditions. The agent works inside that grant and returns with its record — proposed, cited, and staged for your confirmation.

  • Task, scope, and stop conditions declared up front
  • Work staged for review, never applied silently
  • Record kept per assignment for later audit

Consequential steps

A person confirms what matters.

Sending, publishing, deleting, sharing, granting access, and spending always hold for human confirmation. The request is written to be understood in seconds: what is asked, in which workspace, with what scope, and how to undo it.

Try the discipline

You decide. The log remembers.

A front-end demo of the approval moment. Nothing leaves your browser. Tap a verdict and watch the audit trail grow.

Operator asks: send the staged launch notes to the team channel

Workspace billing-demo · scope: one message, one channel · undo: recall within 60 seconds.

Demo only. Real approvals are enforced by the control plane, not by this page.

    Permission model

    Scoped roles, attributable actions.

    The model is deliberately narrow. Read the philosophy for the control-first reasoning behind it.

    • Role-scoped permissionsEach assignment grants the minimum files, tools, network, and time its role needs — bounded per task, never carried over silently to the next one.
    • Every action attributableActions are logged to the acting assignment as they occur: who authorized it, what was done, in which workspace, and under which grant.
    • Human confirms consequential stepsAnything that sends, publishes, deletes, shares, or spends waits for explicit approval. Denials are recorded alongside grants.
    • Stoppable and revocablePause or revoke any assignment from its record. Revocation ends the grant immediately; the record remains for review.

    Control commitments

    No unlogged agents, ever.

    This is the suite's standing refusal, applied to every assignment. Read the philosophy for the full list of what we will not build.

    • 01

      Human in command

      Agents propose, draft, gather, and organize. A person approves anything consequential — and can redirect or stop the work at any point.

    • 02

      Contained by design

      Single-active workspaces with permission jails. No silent scope expansion: anything outside the grant needs a new approval.

    • 03

      Auditable always

      Grants, actions, approvals, and denials recorded in readable language, so any outcome can be reconstructed and reviewed.

    Begin

    Delegate the task, keep the authority.

    Request access to ALPHA Agents, or enter ALPHA Studio today to work inside the same discipline.

    Request access Enter ALPHA Studio