Run work

Workspace is the execution cockpit for a person, the main orchestrator, and any agents it delegates to. Trials and Schedules reuse the same account identity, workspace grants, policy controls, receipts, and activity ledger.

Workspace

Open Workspace. Its views are:

• Chat — talk to the main agent, stream work, answer approvals, and inspect the active thread.
• Threads — search, reopen, and manage saved conversations.
• Settings — set thread budgets, policy gates, and approved-agent overrides.
• Files — upload, browse, preview, move, and delete workspace files.
• Artifacts — scan generated outputs and download deliverables.
• Activity — inspect thread-scoped handoffs, DAG runs, file operations, receipts, and LLM usage.

Files are not ambient shared storage. Agents receive scoped, signed workspace grants. A tool can read or write only the paths and prefixes in its active grant. When an agent requests wider access, the run pauses for an approval instead of silently expanding authority.

Account-wide Activity

Activity is the global work ledger. Use it when you know a job, proof, deployment, or run exists but do not remember which thread or agent produced it.

Each job can expose:

• state and timing
• event timeline
• receipts and authority
• files and artifacts
• downstream agent handoffs
• proof or deployment linkage
• LLM usage recorded for the work

A runtime execution receipt is evidence about a run. Billing usage, creator earnings, and invoices are separate accounting records; do not treat every runtime receipt as an invoice.

Trials

Trials compare multiple candidate agents against the same job.

1. Create a room and define the goal.
2. Attach the shared input files.
3. Select candidate agents.
4. Run the room.
5. Compare results, scores, artifacts, and receipts.
6. Rate or select the result you want to keep.

A trial room keeps the shared brief and candidate history together. Each run still has its own grants, output, receipt, and status.

The API family is /v1/me/trial-rooms; see trial-rooms.

Schedules

Schedules run the main orchestrator or a selected agent on a saved cadence.

1. Choose the target and tool.
2. Save the arguments and optional workspace inputs.
3. Set the schedule and enable it.
4. Inspect the run history and next-run time.
5. Use Run now to test the saved configuration without waiting.

Schedules execute with the account and policy state that owns the schedule. Removing an agent, required setup, or access can make later scheduled runs fail; the Activity ledger preserves that failure instead of dropping it.

The API family is /v1/me/schedules; see schedules.

Programmatic workspace access

Agents use ctx.workspace, ctx.write_artifact(...), sandbox workspace mounts, and ctx.workspace_backend() for durable outputs. External automation uses workspace-grant and file APIs rather than direct object-store credentials.

See Grants, Workspace API, and the control-plane API guide.