Agents (calling a model like anything else you declared)

Summary#

An agent is something you declare, and calling it is calling a name. Its declaration supplies its instructions, its tools and its model, so the call site stays about what you are asking rather than about how the model is configured.

using Osyrin.Agents;

[Principal] entity User {
  [Required, MaxLength(100)] string DisplayName;
  security { allow read when IsAuthenticated; allow create when IsAuthenticated || IsAnonymous; }
}

agent Support {
  Purpose   = "Answer questions about an order.";
  Prompt    = "You answer order questions from the context you are given.";
  Principal = new User { DisplayName = "Support" };
}

string Answer(string question) {
  var turns = new List<Turn>();
  turns.Add(Turn.User(question));
  return Support.Ask(turns).Text;
}

Description#

You hand it the turns it should see. There is no hidden session: the question, the history worth replaying and the context you assembled are the argument (running an agent from your code). What it "remembers" is what you passed — see agent conversation memory (using Osyrin.Agents) for the shapes that are worth passing and the ones that are not.

Every run is a task, recorded. Cost, duration and outcome are captured whether or not you look (the agent task log (AgentTask)). watching a task run (task.Watch) follows one in flight, what the agent saw (task.Transcript) reads what it actually said and did, what a task cost, and what it did (task.Calls) lists the tools it reached for, and stopping work (task.Stop) ends one.

The model is a declaration, not a call-site argument. app.Models — the models an app admits, each with a name is how a model is named and default LLM model (app.DefaultModel) what a call gets when it names none — so changing which model an agent uses is an edit in one place, not a sweep of call sites.

A long-running agent is a workflow, not a loop you write. the agent loop (app.Agent, Loop) is the shape for repeated turns, and an agent asking a person (the human slot) is how a person is brought into one — which is a parking point, with everything the workflow area says about surviving a deploy. what an agent hands back (AgentDeliverable) is how a run's output becomes something the app holds.

The pages#

Run osy docs agent for the full listing.

• Calling one — running an agent from your code, agent conversation memory (using Osyrin.Agents), the agent loop (app.Agent, Loop)
• Which model — app.Models — the models an app admits, each with a name, default LLM model (app.DefaultModel)
• People in the loop — an agent asking a person (the human slot), what an agent hands back (AgentDeliverable)
• Watching a run — the agent task log (AgentTask), watching a task run (task.Watch), what the agent saw (task.Transcript), what a task cost, and what it did (task.Calls), stopping work (task.Stop)

See also#

• running an agent from your code — the call, and what a turn list is for
• the agent task log (AgentTask) — what every run records, without being asked
• an agent asking a person (the human slot) — bringing a person in, and why that makes it a parking point