Taking on new work
Research · 2026

Reaper Desk — multi-agent trading floorActive

A paper-trading research desk staffed by named agents: an analyst drops opportunity briefs, a statistician blends a deterministic score with an LLM verdict, an auditor approves, an executor manages exits. Humans hold the only switch to live money.

Reaper started as a single-signal options bot. Reaper Desk is what it became: a small team of agents with separated duties, built to answer a research question. Can a language model add judgment to a quantitative pipeline without being allowed to act on its own?

The desk

  • Iris (analyst) publishes JSON opportunity briefs from her own container.
  • Moira (statistician) computes a deterministic quant score, then asks a cloud model for a reasoning verdict. If the model is unreachable, she degrades to pure quant and says so.
  • Atropos (approval and audit) independently checks every proposed trade against limits.
  • Charon (execution) places and manages exits.
  • Hestia (portfolio steward) runs a separate, buy-only allocation with drift tracking.

Hand-offs are files on a shared bus, which is how Athena renders the desk as a live diagram with firing edges between agents.

Safety posture

The desk trades paper by default. Arming live mode is a manual, deliberate step outside the agents' reach, and the bot's heartbeat, mode, and errors are surfaced on the Athena status board so a stalled or misbehaving process is visible within minutes. Strategy internals are not published; the orchestration pattern is.

At a glance

StatusActive
TypeResearch
Year2026

AI

Ollama CloudMulti-agent orchestrationGraceful degradation

Stack

Python 3.13Alpaca APIFile-based event busmacOS launchd
Related work