A trading firm where the LLMs can't approve their own trades

What it does

ALTA runs a virtual research firm on your machine: four specialist “Scout” agents hunt for market opportunities (changes, dislocations, causal effects, expectation gaps), independent reviewers challenge each thesis, and a strategy desk compares instruments before anything happens. Execution is split into a Shadow simulation mode and a separate Broker API boundary that requires a verified, explicitly selected account. Every handoff leaves an inspectable record, so you can trace a trade idea back to its source citations.

The interesting bit

The architecture is built around distrust of the agents themselves. Assessors commit their views before seeing each other’s, ranking is done by code rather than a voting model, and a persuasive thesis structurally cannot approve its own execution. Only structured, versioned artifacts cross handoffs — no ever-growing shared chat log.

Key highlights

• Four Scouts with distinct “Trader Minds” — research styles and accumulated lessons — sharing tools but not an agenda
• Execution gates enforced in code: freshness, risk, account isolation, durable transitions
• Operator console with agent desk, evidence views, replay ribbon and a distinct Shadow book vs. broker account view
• Models configurable by role, with separate choices required for opposing assessments
• Runs locally; saved credentials never grant trading permission

Caveats

• Explicitly experimental research software — not investment advice, not a production OMS, and no evidence of profitable alpha
• Six broker connectors exist, but live-account acceptance is not established
• Heavy local stack: Node 22+, Python 3.12, Docker, and a Rust toolchain for the pinned custom harness

Verdict

Worth a look if you’re researching multi-agent governance, adversarial review pipelines, or LLM decision-making with hard execution boundaries. Skip it if you want a trading bot that makes money — the README is refreshingly upfront that it doesn’t claim to.