Tracking: Migrate Agent layer to OpenAI Agents SDK

Inizia leggendo l’architettura target e le PR di riferimento #61, #65, #67 e #70, quindi esamina le aree attuali brain/, tools/, flow/ e llm/. Questa issue è un tracker della migrazione, non un’attività autonoma; è completata solo quando il lavoro di migrazione della roadmap è stato terminato.

Motivation

QuantMind invested ~3 months building a smolagents-derived Agent runtime in PRs #61, #65, and #67 (~1900 lines of brain/ + tools/ + supporting modules). We now believe this was a wrong direction:

• Agent runtime is a commodity in 2026. OpenAI Agents SDK, smolagents, LangGraph, etc. compete on the same surface. Maintaining our own gives QuantMind no differentiation.
• It consumes engineering time that should go into financial-domain work (data schemas, e2e pipelines, research-friendly memory patterns).

This issue tracks the migration to OpenAI Agents SDK as the underlying agent runtime, with QuantMind repositioning as a domain library on top.

Architecture update (2026-07-15): Durable canonical collections and semantic indexing belong to the top-level quantmind/library/ package tracked in #111. mind/ owns cognition, tools, working memory, and later tree navigation; it may consume library but does not own persistence or index lifecycle. This supersedes the vague KnowledgeStore placement in the original roadmap below.

Target Architecture

quantmind/
├── flows/        # e2e processing pipelines (paper_flow, news_flow, ...)
├── knowledge/    # Pydantic-based knowledge schema standard (Paper, News, ...)
├── library/      # canonical collections + derived semantic index (#111)
├── preprocess/   # fetch (arxiv/http/doi/local) + format (pdf/html/markdown)
├── mind/         # cognition, tools, working memory, tree navigation; no persistence
├── configs/      # centralized BaseFlowCfg + per-flow cfg + input types
├── magic.py      # resolve_magic_input: natural language → (input, cfg)
└── utils/        # logger only

Key principle: QuantMind does NOT rebuild Agent runtime, lifecycle hooks, tracing, multi-agent handoff, or tool framework. Those come from openai-agents.

Roadmap

• PR1 (#70): Clean removal of self-built agent runtime
• PR2: knowledge/ + configs/ skeleton (Pydantic schemas + centralized cfg)
• PR3: preprocess/ two-layer structure (fetch + format)
• PR4: flows/ + paper_flow + batch_run + magic.py; concurrent removal of flow/ and llm/
• PR5: mind/memory/filesystem MVP + trajectory archive
• PR6+: Semantic library (#111), paper-search integration (#112), PageIndex navigation (#95), additional sources, and observability cookbook

Archive

The pre-pivot agent runtime is preserved as a snapshot on the archive/agent-runtime-final branch on origin for historical reference. It will not be merged back to master.

References

• OpenAI Agents SDK docs: https://openai.github.io/openai-agents-python/
• Lifecycle / RunHooks: https://openai.github.io/openai-agents-python/ref/lifecycle/
• MCP integration (filesystem server): https://openai.github.io/openai-agents-python/mcp/
• Tracing: https://openai.github.io/openai-agents-python/tracing/
• SDK announcement: https://openai.com/index/the-next-evolution-of-the-agents-sdk/