The infrastructure layer
for OKX's AI and
agent products

OKX's "Onchain OS" and Web3 wallet position AI agents as a core product direction — agents that autonomously monitor portfolios, execute DeFi strategies, bridge assets across 130+ chains, and respond to market conditions without user intervention. Building agents that are reliable at OKX's user scale requires infrastructure that handles the hard parts: persistent state, tool execution, memory across sessions, and graceful failure recovery.

Cloudflare's Agents SDK is purpose-built for this. It handles the orchestration layer — LLM calls, tool invocations, state management, and human-in-the-loop checkpoints — as a managed runtime that runs globally on Cloudflare's network. One agent per user session, co-located with the user.

• Durable Objects for per-user agent state — each OKX user's AI agent is a Durable Object. It holds the agent's memory: portfolio context, active DeFi positions, pending transactions across chains, risk parameters, and conversation history. When the user returns the next day, the agent remembers everything — no stateless cold start, no round-trip to a central session store.
• Human-in-the-loop checkpoints — for high-value actions (large swaps, cross-chain bridges, DeFi strategy execution), agents can pause and request user confirmation via the OKX app before proceeding. Cloudflare's Agents SDK has native support for durable human-in-the-loop pauses — the agent waits indefinitely for confirmation without consuming resources.
• Workflows for multi-step onchain execution — a DeFi rebalancing operation might span: check prices → calculate optimal routes → execute swap on chain A → bridge to chain B → deposit into yield protocol. Cloudflare Workflows makes each step durable and retryable — if the bridge step fails, it retries from the bridge, not from the beginning.
• Browser Run for Web3 agent interactions — agents that need to interact with DApps that don't have APIs (new protocol frontends, NFT platforms, governance portals) can use Browser Run to operate headless browsers at the edge — no browser infrastructure to manage, sessions persisted via Durable Objects for continuity.

Workers AI runs 50+ open-source ML models on serverless GPUs across Cloudflare's global network — co-located with OKX's users. For AI features where latency is critical and frontier model quality isn't required, Workers AI delivers sub-50ms inference from a PoP in Singapore, Tokyo, or Dubai — not from a centralized US cloud endpoint.

OKX's AI models need training data — historical market data, transaction records, onchain signals, model checkpoints, and fine-tuning datasets. At exchange scale, this is petabyte-class data that the AI/ML pipeline reads repeatedly during training runs, model evaluation, and feature engineering. S3 egress charges on this data compound into a significant monthly cost for any team running frequent training iterations.

• $0 egress on training data reads — AI training pipelines read the same historical market data hundreds of times across experiments. R2 eliminates egress charges entirely — reading a 1TB market data snapshot 50 times during hyperparameter tuning is free instead of $50K+ in S3 bandwidth.
• Model artifact storage — fine-tuned model weights, ONNX exports, and inference-optimized model files stored in R2 and served directly to Workers AI for edge deployment — with zero egress when models are loaded for inference.
• Prompt and completion logging for AI Gateway — AI Gateway can log all LLM requests and responses to R2 for model evaluation, fine-tuning dataset generation, and compliance. This is the training flywheel: production inference data feeds back into model improvement without a separate data pipeline.