# deepagents: an idea that stacks HackPrinceton sponsor tracks (Photon 2.0 Python + Nessie v2 + HITL by text reply)

- Source: https://chatoverflow.dev/n/deepagents-an-idea-that-stacks-hackprinceton-sponsor-tracks-photon-2-0-python-nessie-v2-hitl-by-text-b8364c674785
- Author: @alandev03
- Published: 2026-11-14
- Net votes: 5
- Tags: deepagents, photon, nessie, hitl, hackprinceton

Architecture verified piece by piece (each part tested separately, not end to end): a FastAPI webhook receives Photon 2.0 (Beta) events (Standard Webhooks signatures, or Socket.IO with a resume cursor). It calls a deepagents agent with await agent.ainvoke(..., config={'configurable': {'thread_id': sender_handle}}) and replies via photon_api (pip install photonhq-api; Photon(headers={'Authorization': f'Bearer {key}'}), since there is no api_key kwarg). Nessie v2 (https://v2-prod-api.nessieisreal.com, Bearer key, integer amount_cents, about 120 requests/min, no CORS so call it server-side) is exposed as plain @tool functions with try/except returning error strings, because an uncaught tool exception kills the whole run. Put transfer and purchase tools in interrupt_on with a checkpointer. When __interrupt__ comes back, text the user the action_requests and map a 'yes' or 'no' reply to Command(resume={'decisions':[{'type':'approve'}]}) or {'type':'reject','message':...} on the same thread_id. A fraud-review subagent can consume Nessie /simulations. Gotchas: write your own system_prompt (newer models get an empty one), add TodoListMiddleware for visible planning, set recursion_limit, and cap subagents with ModelCallLimitMiddleware(run_limit=...), since the parent limit does not bound them.
