Multi-Agent Crypto Trading

One trading bot acting alone gets wiped out by every news cycle. Built a 4-AI committee for crypto trading. Each AI looks at the trade from a different angle, the bot only acts when they all agree.

• 01User surface

  The whole experience the user sees

  This frame shows the live product: multiple ai agents trade the top 50 crypto tokens together · data, sentiment, news in one. Every component is ours · scope, design, code, deploy.

• 02Stack behind the screen

  What's powering it: Python, pandas, NumPy

  7 stack components run behind this frame · Python, pandas, NumPy drive the visible UI; the rest sit in the data layer. All studio-owned.

• 03What we shipped

  Multiple AI agents for one decision

  Decisions cross-checked from multiple angles

• 04Status

  Open-source reference.

  Full implementation on GitHub · forkable by anyone, auditable end-to-end.

• 01 · BRIEF

  Why a single bot blows up.

  Single-signal bots ride on price-only logic; they get destroyed on news or sentiment swings. Spec'd a quorum-based fabric: at least 3-of-4 agents must agree before a trade fires.

• 02 · ARCHITECTURE

  Stack decisions before any code.

  Decision doc captured the data flow, Python, pandas, NumPy, scikit-learn role split, and the failure modes we'd handle in v1 vs defer. Cross-service boundaries (where AI ends and the web app begins) were drawn here so neither side leaked into the other later.

• 02 · BUILD

  Four agents on LangChain + Ollama.

  Local LLMs via Ollama for the sentiment and verification agents · cost stays bounded. Quant agent uses pandas + scikit-learn on the historical SQLite store. Decisions logged with rationale.

• 04 · POLISH

  Performance, accessibility, and observability.

  PSI / a11y / coverage budgets enforced as launch gates. Logging + metrics wired before cut-over · the team can answer 'is it working?' from a dashboard, not a Slack thread. Threat-model checklist signed off before traffic hits the box.

• 03 · SHIP

  Live · web UI + API + open repo.

  Web UI for human supervision, JSON API for downstream automation, public repo so other studios can fork the agent fabric.

• 01Quorum

  3-of-4 consensus protocol

  No trade fires until at least three agents agree · single-source false positives can't act on their own.

• 02Quant

  scikit-learn model on price + volume

  Pandas pipeline reads the SQLite store, scikit-learn returns the quant agent's vote · explainable.

• 03Sentiment

  Local LLM reads social signal

  Ollama hosts a sentiment classifier locally · zero external API cost per check.

• 04Verifier

  Last-mile sanity-check agent

  Independent agent re-evaluates the trade against the policy before placement · catches the rare consensus-but-stupid case.

• 
  01Frame

  Boot script · run.sh + venv + module imports

  Single-command entry: `./run.sh` (left pane) creates the venv if missing, hashes the requirements file, activates, then boots MACATS · 'importing asyncio core / market data stack (pandas / numpy / yfinance / ccxt) / orchestrator + agent graph'.

• 
  02Frame

  Agents online · 5 spawned, live loop entered

  Boot completes in 1.59s. Agents spawn one per line: FSVZOScannerAgent, RiskAgent, ExecutionAgent, StopAgent, PortfolioAgent · `[ready] all agents running, entering live loop`. Each line is a real Python process you can attach to.

• 
  03Frame

  Live scan · per-token signals JSON

  Strategy log per token (ADA/USDT, BNB/USDT) with the full payload: VWAP drift, score, price, and the F/S/V/Z/O signal map · post-mortem reads the same log, no extra instrumentation.

• 
  04Frame

  Continuous loop · scanner tick #3, queue depth 9

  Scanner ticks every 29 ms with `queue depth 9` visible · AVAX/USDT scan output streams in alongside the others. The system is always-on; nothing waits for a human to refresh a dashboard.