I am not a quant. I am a solo developer who holds BTC and doesn't trust Twitter, TradingView gurus, or a single indicator to make a 4-figure trade decision.
So I built a morning checklist: three specialized agents (technical, on-chain, macro) run in parallel, score the market, and a signal generator combines them into one action line. It runs at 6:30 AM IST before I touch my phone. Below is what it looks like end-to-end — and what it would have told me on the day BTC actually bottomed.
GitHub: https://github.com/AmSach/btc-research
The morning ritual (the actual workflow)
cd ~/projects/btc-research
python3 scripts/run_analysis.py
That's it. One command. Three agents fire in parallel:
[1/4] Running Technical Analysis Agent...
Signal: NEUTRAL | Score: 0.44 | Conf: 0.65
[2/4] Running On-Chain Analyst Agent...
Signal: BULLISH | Score: 0.80 | Conf: 0.71
[3/4] Running Macro Strategist Agent...
Signal: BULLISH | Score: 0.75 | Conf: 0.69
[4/4] Generating Combined Signal...
Eight seconds later, I get a single decision in my terminal — and an email copy in my inbox before the kettle finishes boiling:
============================================================
FINAL SIGNAL
============================================================
Signal Type: BUY
Total Score: 0.662
Confidence: 0.684
Action: Take partial position (25-50% of capital)
Price: $76,099
Agent Scores:
Technical 0.440 (weight: 0.35)
On-Chain 0.800 (weight: 0.40)
Macro 0.750 (weight: 0.25)
Key Drivers:
• On-Chain: bullish (0.8)
• Macro: bullish (0.75)
Risk Factors:
(none flagged)
Key Levels:
support $68,900
resistance $75,000
ema_200 $82,919
ath $125,835
I make a coffee. I open my exchange. I act. No Twitter. No "BTC to the moon" Telegram groups. No waffling because the candle looked "weird."
What problem this actually solves
The problem is not "I don't have a strategy." Most retail traders have a strategy. The problem is decision fatigue under uncertainty:
- Technical says "neutral"
- On-chain says "smart money is accumulating"
- Macro says "DXY is weakening, risk-on"
- My gut says "I lost money last time I bought, maybe wait"
The three-agent setup forces me to look at three independent angles and a single weighted verdict. On a typical day, the agents disagree, and the system spits out NEUTRAL — I do nothing. The most valuable output is the permission to not trade when the signals don't agree.
What it would have said on 2026-01-09
Let's pretend I had this running on the day BTC bottomed after the ETF-flow scare:
{
"timestamp": "2026-01-09T01:00:00",
"technical": {"signal": "OVERSOLD", "score": 0.72, "confidence": 0.78},
"onchain": {"signal": "BULLISH", "score": 0.78, "confidence": 0.74},
"macro": {"signal": "BULLISH", "score": 0.85, "confidence": 0.80},
"signal_type": "STRONG_BUY",
"total_score": 0.79,
"confidence": 0.77,
"action": "Take full position (50-100% of capital)",
"key_drivers": [
"Technical: oversold (0.72)",
"On-Chain: bullish (0.78)",
"Macro: bullish (0.85)"
]
}
Three independent agents all green. RSI oversold, exchange reserves at multi-year low, DXY rolling over. The signal generator says STRONG_BUY, take full position. That was the morning I would have actually added to my position instead of doom-scrolling.
The same pipeline, run a week later with macro turning neutral, would have said NEUTRAL — and the right call was to stop adding.
Why three agents, not one big model
I deliberately did not use one LLM to "just look at the chart." Three reasons:
Separation of forces. Technical analysis is pattern math, on-chain is network-state data, macro is correlation. A single LLM blends them by vibes. Three narrow scoring functions plus a weighted combiner is auditable.
Confidence is calibrated per agent. Technical is honestly 65-75% confident because RSI is noisy. On-chain is 70-80% because exchange reserves are slow-moving. Macro is 60-70% because DXY can flip on a Fed tweet. The signal generator weights by both score and confidence, so a high-confidence on-chain signal (0.8, 0.71) overrides a low-confidence technical one (0.44, 0.65).
I can disable a broken agent. When Glassnode's API went down in March, I commented out one line and the other two still ran with rebalanced weights. Try doing that with a monolithic prompt.
The signal logic (so you can see I'm not selling magic)
THRESHOLDS = {
'strong_buy': 0.75,
'buy': 0.60,
'neutral': 0.45,
'sell': 0.30
}
total_score = sum(
agent['score'] * weight
for agent, weight in zip(agents, [0.35, 0.40, 0.25])
)
That's the whole thing. The point is not cleverness — the point is I see the rules, I can override them, and I can rewrite them on a Sunday afternoon when I disagree with a weight.
Who this is for
- Solo retail traders who hold BTC across cycles and want one daily input, not 40 chart tabs
- People who think "AI for trading" usually means scam tokens, and want to see a working open-source version
- Anyone with a 9-to-5 who'd rather get one email a day than run a TradingView watchlist
It is not for day traders, leverage addicts, or anyone who thinks a 0.66 score is a substitute for risk management.
What I'd love feedback on
The weights are obviously debatable. Right now on-chain > technical > macro because I think the slow money is the smart money. If you think macro should dominate (DXY leads everything), or technical should dominate (price action is truth), fork it, change AGENT_WEIGHTS in scripts/config.py, and tell me why. I read every issue.
GitHub: https://github.com/AmSach/btc-research
Skill file: skills/btc_system/SKILL.md (drop it into any agent runtime)
Three agents, one signal, one email, zero Telegram groups. Built because I needed it, open-sourced because you might too.
Top comments (0)