This is a scoped research dossier, not a completed systematic review or an independently tested result. It identifies methods, questions and source trails for future reporting.
Single-loop architectures
Repeated model → tool → observation cycles can be powerful. Each cycle needs termination limits, error handling, identity and authorization checks.
Orchestration patterns
Supervisor/worker and graph-based approaches differ in observability, handoff state and failure propagation. More components do not automatically mean better results.
Durability matters
Long-running tasks must survive interruptions without repeating expensive or destructive steps. Idempotency and checkpoints deserve as much attention as prompt quality.
What would count as evidence?
Compare an identical task on a single-agent loop and controlled multi-step graph.
Documents to examine
- OpenAI Agents SDK — Running Agents
- MCP 2026 Tools Specification
These are starting points, not claims that every document has been independently reproduced.
Read our cited field note →Edition 1.0 · 09 October 2026
Initial research brief published. No earlier revisions or submitted public corrections are claimed.
Suggest a documented correction ↗