Agent Studio · Build
An agent is a job description, not code.
Agent Studio is where the agents of AgenticObjects’ agentic analytics are written: what an agent watches, for whom, within which limits and on which schedule. An agent stays a draft until every required part is filled, and the screen says which one is missing.
What you write
Nine parts, in the order a manager would brief a new analyst
- Persona Audience, tone, priorities and the language its answers are written in.
- Core identity What is out of scope. A metric not in the model is absent, not proxied. “Moves together with” is the strongest verb: no causation.
- Discovery instruction The signals to watch, the thresholds, the noise rules, and the order of work: base measurement first, then narrow.
- Model note What works in the model and what returns empty, read by every agent bound to it. How far the data runs is measured by the engine, never typed here.
- Data and scope The whole model or one of its Subject Areas. An empty scope is no access.
- Output contract Which record types it publishes and how many of each, at most. A ceiling, not a target.
- Limits Steps and money per run, tighten-only from the workspace baseline, frozen when the run starts.
- Schedule and distribution Daily, weekday, weekly or monthly, written as a sentence in its own timezone. Recipients in the app and by e-mail.
- Serving model Chosen by a person from an approved catalogue. It never changes on its own: no fallback, no quality escalation.
Before it runs
You can read exactly what it will be told
Compiled, never written
The instruction is compiled from fixed layers in a fixed order. The same inputs give the same text and the same hash, and the run record keeps the instruction that ran.
Tools are structural
A tool that is not in the run’s manifest cannot be called. The pair that lets an agent answer or decline cannot be removed, because an agent without them could do neither.
Lessons wait for a person
A lesson an agent proposes from its own scheduled run changes nothing until a person approves it, unless an administrator switches on that agent’s auto-learn.
On a schedule
A scheduled run knows where it stands
Nothing new, nothing spent
When the data has not moved since the last scheduled run, the run is skipped before any model call and costs nothing. After three skips in a row the schedule pauses itself and says why.
Run now, and watch it
A run on demand shows its live activity: what it is doing, how close it is to each limit, what it has spent and which model serves it.
Experiments stay experiments
A person with the permission may raise a run’s budget as an experiment. Its records are kept, never delivered and never counted in quality figures.
Agents start from blank, from a copy of another agent or from one of three small templates. An assistant that drafts an agent from a business description or an industry template is on the roadmap, not in the product.
Write the first job description with us.
In week two of the pilot, the first discovery agent is written together with your team, and its first brief arrives that week.