AI, Compute & Integrations
The AI model never independently decides to move funds. Its job is to turn a natural-language request into a structured plan or tool call. Deterministic application logic then validates that plan (balance, token, limits) before anything executes, and every fund-moving action still needs your explicit confirmation.
Inference architecture
| Layer | Role |
|---|---|
| CapIX | Primary inference route, currently serving meta-llama/llama-3.1-8b-instruct through DeepInfra |
| Self-hosted Ollama | Secondary route, used if CapIX isn't configured |
| OpenRouter | Legacy fallback, kept available if neither of the above is set up |
Every agent (DCA, Volume, Whale Tracking, Hedge Fund, Token Research, Due Diligence, EasyA Analysis, and Wallet Monitoring) shares this same inference layer and the same request/response format, so a change here applies platform-wide rather than agent-by-agent.
Research agents: tool-calling, not free text
The research and monitoring agents (Whale Tracking, Token Research, Due Diligence, Wallet Monitoring, Hedge Fund, EasyA Analysis) don't let the model answer from memory. Each one runs a tool-calling loop: the model requests specific on-chain lookups or calculations, application code executes them against real data, and the model's final reply is built from those results, never invented. If a number isn't available from a tool, the agent reports it as missing rather than guessing.
CapIX compute partnership
CapIX (capex.network) is BITAGENTS' compute partner, supplying the inference backing the primary route above. BITAGENTS is CapIX's first commercial client.
Technical note
Compute needs scale with the agent. A DCA plan is a light, periodic job. The Hedge Fund Agent's 18-analyst synthesis and the research agents' multi-step tool calls run heavier, more frequent inference. Architecture and cost here are agent-specific, not one-size-fits-all.