How AI Agents Manage a Crypto Trading Bot

Most "AI trading bot" landing pages describe a single magic model that picks winners. The reality of running a live crypto trading bot is messier and more interesting: a team of AI agents manages the bot the way a small quant desk would. One agent writes code, another investigates bugs, a third researches new strategies, and a fourth coordinates them. This post explains how that works, what the signal pipeline looks like, and where the honest limits are today.

Why a Single Model Is Not Enough

A trading bot is not one problem. It is five problems stitched together: market analysis, risk enforcement, code maintenance, bug diagnosis, and strategy research. No single prompt handles all five well. A model that is good at reading charts is wasteful at editing Python. A model that is good at code review is too slow for a live signal loop.

The fix is multi-agent specialization. Each agent owns one job, has a short prompt tuned for that job, and hands off to the next agent when the work leaves its scope. SuperKamouBot runs four core agents:

• CEO — coordinates. Reads the operator's goals, dispatches specialists, verifies outcomes, reports back. Does not write code or analyze charts directly.
• coding-agent — the only agent allowed to edit .py files, change runtime config, or deploy to the VPS. All other agents route write requests through it.
• researcher — designs and backtests new strategies in shadow mode before they ever touch live capital.
• investigator — gathers facts. Reads logs, queries the local DB, scans the codebase, and returns compressed summaries so the reasoning agents stay focused.

A fifth role, Big Boss, sets vision and risk boundaries but does not manage day-to-day execution. Think of it as the board, not the CEO.

The Signal Pipeline: T1 → T2 → T3

Agents manage the bot, but the bot itself generates crypto trading signals through a three-stage pipeline. Each stage has a different job and a different failure mode, which is why they are separated.

T1 — Mathematical Ranking

Every cycle, the bot scores all tradable perpetual futures pairs on technical factors: momentum, volatility, relative strength, market structure. This is pure math, fast and deterministic. It does not decide what to trade — it narrows 30+ pairs down to a handful of candidates.

T2 — Rules-Based Veto

Candidates then pass through hard rules. Risk caps, exposure limits, regime filters, side bans. A candidate can be mathematically attractive and still get vetoed because the bot is already at max exposure, or because the regime filter says "this strategy does not trade in this condition." T2 is non-negotiable. No agent and no model can override it.

T3 — AI Refiner

Surviving candidates reach a language model that reads the context — recent price action, order flow, funding rates, the strategy's own history — and either confirms, adjusts sizing, or rejects. This is where "AI" in the literal sense enters the loop. The refiner cannot bypass T2. It can only refine what T2 allowed through.

This pipeline is what we mean by automated crypto trading: a deterministic core (T1 + T2) with a language-model judgment layer (T3) on top, all bounded by hard risk rules.

What the Agents Actually Do Day to Day

The agents are not in the signal loop itself — the loop runs in code on the VPS. The agents manage the bot around the loop:

• Monitoring. When win rate drops or expectancy goes negative, the CEO dispatches the investigator to pull logs and find the cause.
• Fixes. The coding-agent ships a patch on a feature branch, tests it, and deploys via a scripted path that restarts the systemd service.
• Research. The researcher runs new strategy variants in shadow mode — paper trading against live data — and only strategies that pass a statistical gate (N≥30 trades, Sharpe > 0.5, positive expectancy) are even eligible for live promotion. An automated engine, not an agent, decides promotions.
• Governance. Every change is logged. Every trap (a recurring failure mode) gets a post-mortem in a registry so the same bug is not diagnosed twice.

Honest Performance, Today

This is the part most marketing pages skip. Here is the real state as of writing:

• Live balance: ~$242.64. This is a small test account, not a fund.
• Win rate: ~38.9% over the recent window. Below 50%, and we say so.
• 7-day expectancy: negative. We are actively fixing this with a time-stop and hour-exclusion tuning to cut trades in low-edge windows.
• Live strategies: 3. Shadow strategies in testing: 49.

We publish live performance data at superkamoubot.com/results rather than cherry-picking a backtest. If the numbers are bad, the page shows bad numbers. That is the point.

Risk Controls That Cannot Be Bypassed

Two rules are hardcoded and no agent can override them:

1. 1% risk cap. No single trade risks more than 1% of total balance.
2. 15% max exposure. No single position exceeds 15% of total balance.

A third rule governs promotion: shadow before live. No strategy trades real capital until it clears the statistical gate in shadow mode. Agents do not promote strategies by hand — an automated engine does, based on the numbers.

How to Use This

You don't run the bot yourself. You subscribe to the signals it produces.

• Signals API — JWT-authenticated endpoints at api.superkamoubot.com for developers who want to pipe signals into their own system.
• Pro plan — $29/mo — for individual signal access.
• Elite plan — $49/mo — for higher-rate API access and priority support.
• Payments in crypto — USDT, BTC, or ETH.

Read the full architecture at superkamoubot.com/how-it-works, see live numbers at superkamoubot.com/results, and pick a plan at superkamoubot.com/pricing.

The governance framework that coordinates the agents is open-sourced at github.com/kaboumou/ai-agent-governance-framework. Note: that repo covers agent governance, not trading logic. We do not publish the strategy alpha.

The Honest Pitch

If you want a bot that promises guaranteed returns, this is not it. If you want AI-powered crypto trading signals with transparent performance data, a multi-agent team actively fixing what is broken, and 49 strategies being stress-tested in shadow before they ever go live — that is what we are building, and the numbers are public.

Subscribe, watch the results page, and judge for yourself.