Agentic Sessions

An agentic session turns one or more user prompts into bounded model decisions, approved tool calls, retained conversation state, and a final result.

Two execution regimes

flowchart TB
  P[Prompt and allowed tool surface] --> C{Execution regime}
  C -->|adaptive next action| L[LoopAgentSession: decide and execute one step at a time]
  C -->|plan-first dependencies| S[SOPAgenticSession: generate, validate, and run LightSOPLang]
  L --> R[Answer, awaiting input, interruption, or bounded failure]
  S --> R
  classDef input fill:#eef4ff,stroke:#1c4f9c,color:#0f2557;
  classDef decision fill:#fff7e8,stroke:#b7791f,color:#633c00;
  classDef runtime fill:#f5f1ff,stroke:#6657b8,color:#2d246b;
  classDef outcome fill:#eefbf4,stroke:#28875d,color:#174c38;
  class P input; class C decision; class L,S runtime; class R outcome;

Agentic session execution regimes

Loop Sessions suit tasks whose next action depends on each result. SOP Sessions suit orchestration that benefits from an explicit dependency plan, validation, parallel readiness, and bounded replanning.

Shared contract

Both session types use LLMAgent for model calls, enforce a supplied tool or command boundary, support a supervisor, carry model tags and reasoning effort, expose the latest result, and propagate cancellation through abort signals. They retain state in memory for continuation but do not persist themselves across process restarts.