What is AlphaSwarm?
AlphaSwarm is a platform for researching, testing, and running automated trading strategies — with AI agents doing much of the research legwork, and humans holding the keys to anything that touches real money.
This page is written for readers who are new to the platform or non-technical. No code, no jargon without an explanation. If you are an engineer looking for setup instructions, jump to the Quickstart; if you want the full documentation map, see the Documentation index.
The current public alpha is 0.1.0-alpha.1. That is the
deployment version operators should expect on Kubernetes labels and
alphaswarm-cli --version. See Release notes.
The idea in one paragraph
Quantitative trading means using data and statistics — rather than gut feel — to decide what to buy and sell. Doing it well requires running thousands of experiments: testing strategy ideas against years of historical market data, discarding the failures, and carefully promoting the survivors. AlphaSwarm industrializes that loop. Teams of AI agents generate and test strategy ideas around the clock; the platform records every experiment honestly; and a series of safety gates — statistical checks, risk checks, and mandatory human sign-off — stand between a promising idea and real money.
How a strategy travels: the lifecycle
Every strategy on AlphaSwarm moves through the same four stations:
- Research — a human researcher or an AI research agent proposes a strategy: a precise rule for when to buy and sell. Ideas draw on market data, news, filings, and a knowledge graph of financial relationships.
- Backtest — the strategy is replayed against historical market data: "if we had run this for the last five years, what would have happened?" Good backtests are necessary but famously easy to fool — which is why the platform tracks every attempt (see "Honest statistics" under the safety model below).
- Paper trading — the strategy runs against live market data but with pretend money. This catches problems that historical replay can't: slow data, broker quirks, real-world timing.
- Live trading — real money. Crossing into this stage is called promotion, and it is deliberately hard (next section).
The safety model
AlphaSwarm's safety design rests on a few ideas that show up everywhere in these docs:
- Two worlds, one gate. Research tooling (including the AI agents) lives in one world — you'll see it called the research plane or LLM plane. Anything that can move real money lives in another — the money plane. Code in the research world cannot reach the money world directly; the promotion gate is the single doorway, and it is locked from the money-plane side.
- Humans in the loop (HITL). Promotions and unusually large or risky orders wait in an approval queue until a person explicitly approves them — with extra identity checks (step-up MFA), a rule that the approver must be a different person from the requester ("four-eyes"), and, since August 2026, an expiry timer so stale requests lapse instead of firing late.
- Honest statistics. If you test 10,000 random strategies, a few will look brilliant by luck. The platform keeps a trial ledger of every experiment ever run, so the promotion gates can apply statistics that account for how many attempts were made — a lucky one-in-ten-thousand result doesn't pass.
- A kill switch. Operators can halt all trading immediately, and the system halts itself if its own heartbeat stops. See the kill-switch incident response runbook.
- Everything is recorded. Orders, fills, agent decisions, approvals, and configuration changes land in append-only ledgers, so "what happened and who approved it?" always has an answer.
- New features start switched off. Platform changes ship behind feature flags — switches that default to off — so operators choose when new behaviour turns on, per environment.
Who does what
| Role | What they do on AlphaSwarm | Where their docs live |
|---|---|---|
| Researcher / quant | Designs strategies, runs backtests, reviews agent research | Concepts → Strategy & ML, Tutorials |
| Operator | Runs the platform day-to-day: deployments, approvals, incidents | How-to → Operations, Runbooks |
| Administrator | Manages organizations, users, permissions, cloud accounts | Concepts → Identity + tenancy |
| AI agents | Generate ideas, run analyses, monitor markets — inside the guardrails above | Concepts → Agentic |
The pieces you'll hear named
- Bots — the smallest deployable unit: one strategy plus its data, risk limits, and target environment, packaged so it can be backtested, paper-traded, or deployed as one thing. See Bots.
- Agents and crews — AI workers (research analysts, traders, selectors) that collaborate on tasks. See Agents.
- The knowledge base and graph — the platform's memory: documents, research papers, and a graph of financial relationships that agents query. See Knowledge base.
- The data plane — market data flowing in from vendors, stored in governed catalogs, streamed over Kafka. See Market-data vendors and Streaming governance.
- The operator UI — the web application where humans watch runs, approve requests, and manage the platform.
Where to go next
- "I want to see it run." → Quickstart (requires Docker and a terminal) or ask your administrator for access to the hosted demo.
- "I want to understand the architecture." → Architecture.
- "I need to look up a term." → Glossary — start with the Plain-English basics section.
- "What changed recently?" → Recently added and the changelog.