Agents Studio

The home of every agent

Agents Studio, planned for W1, is where a Digital Employee is defined and supervised. An agent definition holds role, mission, responsibilities, an out-of-scope list, class (executive, manager, specialist or utility), a mandatory owner, a supervisor, skills, knowledge scopes, a model policy, a compute budget, a schedule, escalation rules and a persona reference. Agents are created from a Store template version, with template fields locked and tenant fields editable, or drafted from a role description.

Skills are Actions

A skill is an Action of a subscribed module. The skill picker hides Actions the tenant is not entitled to, entitlement is rechecked at every run, and a lapsed module pauses the skill with a reason. Each skill stores its own autonomy level (A0 to A4), maximum risk tier, conditions and limits. Content from emails, web pages and documents can never authorise R3 or R4 Actions. Before R2+ skills that use connector data, connector health is checked and a degraded connection hands off to a human.

Earned autonomy

The autonomy ladder engine promotes and demotes per skill, and every change is audited. In shadow mode (A1) the agent proposes and people act, and agreement is scored per decision and per field. Eval cases are generated from Action examples and tenant golden tasks, with at least 20 per skill for template agents. The Eval Coverage Score across seven dimensions gates promotion. Automatic demotion follows a critical eval failure, an S3+ incident, an undo spike or a repetition loop, and loop, storm and runaway detectors stop a run and open an incident.

Versions and testing

Agent versions are immutable and pin prompts, skills, pack versions, model policy and persona. A test console runs a draft against a sample input or simulator scenario and shows steps, retrieval, tool calls, policy decisions, cost and latency. Canary releases with automatic rollback, version diffs with eval and cost comparison, and YAML export are planned.

Supervision

Each agent has a desk: current queue, in-flight runs, recent decisions, autonomy per skill, budget burn and incidents, with pause, demote and budget controls. If the owner leaves or loses a permission, the agent loses it too and pauses after a grace period. An "asks from agents" inbox lets an agent request an Action it may not take, and the human executes it. Hindi and Tamil working languages are enabled only after per-language evals pass.