Export GIF Settings

60 frames · 20 fps · 3s

Plato Agent Loop Data Flow

What enters the loop, what the loop does, and what comes out

01 / Authority 02 / Context 03 / Loop Core 04 / Tools 05 / Outputs TaskBus Claim PublishedTask authority contract Workspace Facts files, events, ASK state Control Facts policy + interrupt guard Context Manager bounded execution view render Tool Schema allowed actions policy LLM Step reason + choose chat Tool Parser tool_call to action action Loop Transcript assistant + tool msgs next turn Autonomy Gate proceed / defer / skip risk Runtime Tools files, shell, web execute Event / Audit actions + observations evidence Workspace Effect file diff / command mutation ASK / Confirm waiting for user block Task Outcome done / failed / wait TaskBus User Surface activity + result visible task contract intent, instructions, criteria state facts files, recent events control facts ASK, interrupt, policy rendered input bounded LLM messages tool list schemas + finish tool loop memory prior observations tool_calls or final answer action candidate typed BaseAction approved action proceed path defer / ask user decision needed observation tool message for next turn event facts action + observation workspace change file / shell effects finish / error terminal summary projection activity + result card inspectable proof audit / diagnostics Legend primary data policy / PII async batch data store

What enters the loop

  • • TaskBus claim supplies the accepted PublishedTask contract.
  • • Context Manager turns task, workspace, event, ASK, guidance, and policy facts into bounded LLM messages.
  • • Tool schemas define what the model may request; they are not hidden superpowers.

What the loop does

  • • Each step calls the LLM, receives tool_calls or a final answer, parses calls into typed actions, and checks control boundaries.
  • • Runtime tools execute approved actions and return observations.
  • • Observations are appended back into the transcript so the next LLM step reasons over what actually happened.

What leaves the loop

  • • Workspace effects become file changes, command evidence, or web evidence.
  • • ASK and confirmation can pause execution instead of letting the Agent guess.
  • • Event, audit, task outcome, and activity projections make the result inspectable.