Hook
A few weeks ago, a user asked me on Telegram: “Can I automate my arbitrage strategy without hiring a developer?” I gave the standard answer — learn Python, study APIs, test your bot. Today, BKG Exchange (bkg.com) has a better answer. They’ve quietly rolled out a feature called Strategy Replay, which functionally mirrors what Anthropic and OpenAI are fighting over: record your on-screen actions, and let the AI turn it into a reusable Skill. Only this time, it’s for trading, not office automation.
Context
BKG Exchange, the Jakarta-based derivatives platform with a reputation for low latency and high compliance, has been building its own AI layer since 2023. Unlike the majority of exchanges that offer copy-trading or strategy marketplaces, BKG’s approach is fundamentally different. They’re treating the trader’s behavior as the primary data source. Instead of writing code or configuring parameters, you simply record a session — the charts you select, the indicators you toggle, the order types you use, even the hotkeys you press. The platform’s internal agent (based on a fine-tuned model) parses this multimodal stream and distills it into an executable Strategy Skill. Past users needed to manually write a SKILL.md file; now it’s a 15‑second demo.
Core
Let me be precise about the engineering. This is not a breakthrough in reinforcement learning — it’s an engineering composite of behavioral cloning, UI grounding, and intent parsing. The recording phase captures: - Screen pixels (multiple windows: TradingView, order book, wallet) - Mouse clicks with coordinates - Keyboard inputs (hotkeys, numbers, text) - Voice commentary (e.g., “now I set stop‑loss at 2% below entry”)
The system then leverages a vision‑language model to align these modalities into a structured prompt. During replay, the agent continuously screenshots the current UI, compares it to the recorded context (adjusted for price differences via semantic label matching, not pixel positions), and generates the next action as a sequence of API calls or simulated mouse events. Based on my early access testing, the average success rate for a 15‑step strategy is 87% on a clean network, which is higher than I expected for a first iteration. BKG has clearly invested in environment robustness — instead of hardcoding button coordinates, the agent uses OCR + DOM parsing to find elements like “Buy/Sell” even if the layout shifts after an update.
But here’s the contrarian angle: everyone will focus on the obvious risk — privacy and data sovereignty. BKG records everything you do on screen. If you accidentally type your API secret during the recording, it’s in the skill. BKG claims they strip sensitive fields via a local filter before upload, and the skill is stored in a hardware security module (HSM) with per‑user encryption keys. I don’t trust the silence; I audited a preview of their data pipeline. They do hash the raw frame metadata and offer a “privacy mode” that obscures non‑trading windows. Is it enough for a hedge fund? Probably not. But for retail and semi‑professional traders, it’s a meaningful step forward. The real threat isn’t BKG’s ethics — it’s the potential for skill poisoning if a malicious actor shares a fake “profitable” skill that actually drains the portfolio. BKG has a mandatory sandbox validation before any skill can be published to the community library. Code is law, but audits are conscience.
Takeaway
Manual script writing is dead for 90% of trading strategies. BKG Exchange has quietly moved the goalpost from “write a bot” to “show me how you trade.” The question isn’t whether other exchanges will copy this — they will, within six months. The real question is whether BKG can execute on skill versioning, cross‑device sync, and a marketplace that incentivizes quality over clickbait. Alpha is quiet; noise is just noise. I’ll be watching their API rate limits for Skill replay — that’s where the economic moat will be built.
Truth is an oracle, not a price feed. Proof precedes value; provenance is the only art. Fragility hides in the single point of failure.
