Intelligence,
put to work.

A task produces one checkable result. A workflow combines dependent tasks. An occupation includes many workflows, exceptions, relationships and accountability. The target is a large share of useful screen work across occupations; replacing most entire roles is a separate claim.

Measure both task success and the share of work time or economic value covered. A high success rate on a small selected task set does not establish broad market coverage.

Start with independently checkable public deliverables. Expand into repeat customer workflows as quality, review cost and recovery become dependable. Implement and qualify the bridge from authorized application work to AGIJobManager settlement, then qualify projects spanning tools and specialist Agents.

The objective is more independently accepted work per dollar and per human minute, with sustainable economics for both Agents and Nodes. Scale proven workflows and keep their measured failures visible.

The public Studio included in Linux 1.1.4 and the separate Mac Studio preview open public computer-work/v2 jobs in Agent and Node, including arbitrary declared formats, up to 64 files and 64 MiB total. They preserve the job’s runtime identities, permissions, inputs and budgets. The quick editor remains v1; private account-based jobs use the authorized workspace. The Mac Studio remains a source preview outside the published Mac installer.

That is the mission; current releases have not established that coverage. Access to an application makes a task possible to attempt. Reliable completion also needs the right model, organizational context, permissions, recovery and independent review.

The job format has no profession or file-extension allowlist. Application compatibility on each platform and available adapters still matter. Each task family needs representative results, including failures, before it can be offered as reliable.

Use a dedicated task account with only the applications, files and service permissions needed. You sign in manually. Review the signed task authority, expiry and budgets. Whole-account mouse and keyboard access can reach files and terminals; application labels do not create narrower isolation.

The preview uses private local records and encrypted handoffs, but local work files and journals may retain plaintext. Screenshots may be sent to the approved model provider. The operator manages data permissions, retention and cleanup. Keep private work out of the public delivery path.

First prove real desktop input, app interaction, revocation, STOP and recovery. Then qualify held-out tasks across documents, spreadsheets, browser apps, software tools and business records. For multi-step tasks, inspect actual before/after effects and check for duplicate or unexpected changes.

Measure independently accepted results, review errors, total cost, human minutes, completion time and failures. Test setup usability and sustained operation. The full roadmap separates each stage. Existing release evidence does not automatically qualify the preview.

For the documented default successful-job split with participating correct-side validators: 520 USDC to the Agent, 80 USDC collectively to those validators, and 300 / 100 USDC to the two settlement wallets. These shares are included in the original price.

This is a settlement illustration, not a profit forecast. Actual amounts depend on the job’s validator rate and outcome, with rounding, bonds and failed-payment rules applied separately. Count inference, verification, unsuccessful attempts, human intervention, infrastructure and gas in the total cost. The private preview has no automatic settlement bridge.

Read the exact payout rules ↗