Agent-owned execution DAGs

A chat records its messages, model requests and directly dispatched tools. Ordinary Python helpers do not append structural nodes. Calling a different Agent starts an independent DAG; the parent retains only the invocation/result and a child-session reference.

Ownership

Runtime configuration, permission restrictions and cancellation remain inherited. Graph writers, call IDs, forced node IDs, predecessor and context-history anchors are local to the child. Explicit named Context content remains available; parent graph history is not implicitly imported. Returning, raising or cancelling restores the parent bindings. Same-Agent helper methods reuse the owner without producing helper nodes. Registered tool dispatch remains recorded; ordinary internal calls to helper functions do not become tools.

Implementation plan

  1. Add public-entry failures for managed recursive helpers and different-Agent calls under a chat writer.
  2. Separate lightweight helper execution from recorded operations, and introduce an Agent-owned child-session scope using existing stores.
  3. Verify sync/async behavior, nesting, permissions, cancellation, runtime injection and parent restoration; update old tests that required helper nodes.
  4. Run independent specification then quality reviews, full verification, integrate locally and refresh the default installed App.

Acceptance

Helper recursion creates zero new nodes. Parent nodes contain no child internals. Child nodes have no parent-graph caller IDs. Child records remain readable after return. Failure and cancellation restore bindings. No provider or external action is needed for the tests. Existing histories remain unchanged.

Scope

Agents own their execution graphs. A tool deadline records a tool error without cancelling the parent turn. Browser deadline enforcement remains a separate concern. No new database, public decorator, or separate local runtime instance is introduced.

Implementation status

Graph ownership code and public-entry regression tests are present. Independent specification and quality reviews pass. Optional recording diagnostics retain their existing local warnings. Verify the full suite and the installed App separately; source tests do not establish installed behavior.