Skip to main content
A turn converts an observed message into bounded agent work and, when appropriate, a reply. Formation records what happened before using it as the basis for later work.

​
From message to run

  1. Authenticate and route. The connector verifies the inbound request and identifies its configured agent install.
  2. Normalize and capture. Connector-specific messages become attributed events. Slack, Telegram, and phone connectors commit intake before acknowledging delivery; see durable intake for the connector-specific guarantees.
  3. Decide whether to engage. The addressing gate distinguishes a message the agent should answer from one it merely observes. A name candidate may need bounded classification after durable capture.
  4. Admit work. The runtime checks current agent/install state, conversation ownership, and capacity. Ready triggers can coalesce; their messages remain in the conversation record.
  5. Resolve context. The engine loads this agent’s persona, authorized history and memories, tools, and applicable continuation state. It applies audience checks and prompt budgets.
  6. Execute. The harness records model calls and tool outcomes, applies live corrections, and advances its durable checkpoints.
  7. Finalize. The engine saves the completed turn and rechecks the destination audience and source permissions before delivery.
Identity-linking and memory-sharing commands are authenticated runtime operations; a model cannot approve them by interpreting a message as consent.

​
From model answer to delivered reply

Model text is canonical content. Connector rendering adapts it for the destination at send time: for example, Slack formatting and attribution are presentation, not edits to the canonical answer. A durable reply boundary prevents recovery from blindly sending the same result again. The actual connector message ID is available only after sending; Formation then records the outbound message event and run outcome. A failure between those steps can leave an uncertain delivery. It must be reconciled, not treated as permission to repeat inference or send another copy. Replies retain message, memory, and supported tool-source provenance so later history retrieval cannot strip the restrictions on their original evidence.

​
Work that outlives one request

Live steering updates an active conversational run. Continuation restores a paused task after a new requester reply. Checkpoint recovery resumes an interrupted run under its existing durable budget and current authority. These are different lifecycles; restarting a worker is not new human consent. Commitments and schedules have their own durable triggers and authorization. They use the execution boundaries above without fabricating a new human message. General WorkGraph execution is not implemented. Implementation: internal/app (repository), internal/graph (repository), and internal/turn (repository).