Docs
GitHubBlog

Search Documentation

Search for a page in the docs

AutoQuant

AutoQuant is OpenAlice's Agent-native quantitative research activity. It uses one durable Workspace backed by an exact release of AutoQuant V2, while OpenAlice supplies the native Agent Session, collaboration, Inbox, market-data tools, and optional UTA access around that research desk.

OpenAlice does not reproduce AutoQuant's Project, Study, Research Session, Run, Report, or Dossier lifecycle. The coding agent works inside the AutoQuant repository and follows its own AGENTS.md.

Initialize the desk

Open Quant from the main activity navigation.

Quant uses the same responsive navigation shell as Ask Alice: focused, recent, and tree views behave consistently, while AutoQuant keeps its own copy, readiness gate, and display preference. There is no separate AutoQuant frontend shell or headless UI.

  • On a fresh installation with no AutoQuant V2 Workspace, initialize one durable desk before sending research.
  • If compatible desks already exist, select the one that should become the default; OpenAlice does not guess.
  • Later assignments create or continue Sessions inside that selected desk rather than creating another Workspace.
  • The low-frequency Workspace control lets you change the default desk explicitly.
  • Browsing an older Session does not silently make its Workspace the durable AutoQuant default. Only explicit Workspace selection or creation changes the default, and OpenAlice saves that choice before navigating.

The default source is AutoQuant V2 v0.9.31 at commit adc6363a7af5a9105811735973d4d5cfac58cf36. The creation flow also retains approved v0.8.31, v0.8.30, and v0.8.27 sources for reproducible older desks. Versions are exact creation inputs, not floating branches or ranges.

Delegate from Ask Alice

New or explicitly template-upgraded Chat Workspaces receive a delegate-autoquant skill that teaches the Chat Agent to hand a bounded quantitative research assignment to the initialized default desk. The underlying public route is:

alice-workspace conversation ask --harness autoquant \
  --prompt "Research this hypothesis, preserve the stated constraints, and return the evidence paths"

Add --await when the current answer needs the result. Otherwise retain the returned taskId and later use conversation await, read, or collect. AutoQuant delegation has no default execution deadline; use --timeout-ms only for a caller-chosen hard stop.

The Harness target requires an explicitly initialized default AutoQuant Workspace. It never creates or guesses a desk during delegation. The returned Agent message should name its Project and applicable immutable evidence ids, give the absolute Project root and primary deliverable directory, distinguish validation from holdout evidence, and state assumptions and the no-trading boundary. Continue the returned resumeId for clarification instead of recruiting another worker and losing its context.

AutoQuant does not automatically publish a Report or Dossier to Inbox. Its ordinary Agent reply is the handoff; request inbox push only when a durable human-facing delivery or asynchronous notification is actually wanted.

Source and provenance

Bootstrap retains the upstream Git history and canonical origin, creates a fresh research branch at the selected commit, and records the exact source in:

.alice/harness-source.json

The Workspace keeps the upstream AutoQuant AGENTS.md. OpenAlice does not replace it with the Chat template's Alice instructions; it adds discoverable collaboration, market-data, Inbox, and UTA skills alongside the upstream instructions.

Python 3.11, uv, AutoQuant dependencies, Projects, Studies, and research iteration remain the coding agent's responsibility inside the desk. A useful first orientation is:

uv run aq project list .
uv run aq validate .
uv run aq orient . --json

Boundaries

  • AutoQuant owns quantitative research truth and evidence.
  • OpenAlice owns Workspace, Agent, Session, collaboration, and delivery lifecycle.
  • UTA alone owns live accounts, approvals, and trading writes.
  • Existing Auto-Quant Classic Workspaces are left intact and are not reinterpreted as V2 desks.
  • Source versions are never upgraded automatically. Generic Template Upgrade does not apply; a future AutoQuant-aware upgrade must preserve upstream history and explicit source choice.

Next Steps

  • Workspaces — Understand the durable desk and Session model.
  • Templates — See the source-aware auto-quant-v2 template.
  • Inbox — Publish evidence-bound reports back to the user.
  • AI Providers — Configure the coding Agent used by the desk.