Most AI agents today are retrieval-augmented or tool-calling wrappers around LLMs. They fetch context, format prompts, and generate text. They cannot reason causally — they cannot discover what causes what, compute interventions, or perform counterfactual analysis. Building an agent that can autonomously perform causal analysis requires a fundamentally different architecture. This post outlines how Abel's causal agent is built.
The core loop is graph-walk plus evidence-search. Given a user question ("Should I all-in AI content creation?"), the agent (1) maps the question to anchors in the causal graph — nodes like NVDA, SNAP, BILI, GOOG — (2) runs causal discovery to find parents, children, and propagation delays (tau), (3) walks the graph to trace causal chains, (4) runs evidence search to validate or refute hypotheses with dated catalysts, (5) iterates until the causal picture converges or confidence saturates. This is not a single prompt — it's a multi-step computational workflow where each step informs the next.
The agent must distinguish correlation from causation. It uses PCMCI and 38 other algorithms for conditional independence testing, Granger causality, and structural discovery. When the graph says "CCOI → SNAP with tau=35h, weight=-0.099," that's a causal claim, not a correlation. The agent then searches for mechanistic evidence: "Why would internet backbone (CCOI) cause content platform (SNAP) pressure?" Evidence like "AI wavelength revenue +150% YoY, fiber being consumed by AI training" validates the edge. Without this, the agent would output spurious associations.
Intervention and counterfactual primitives complete the loop. Once the graph is discovered and validated, the agent applies do-calculus to answer "If I do X, what happens to Y?" and, in beta, counterfactual reasoning for "If X had been different, would Y have changed?" The output is not a narrative — it's an Insight Card: direction (bearish/bullish), key causal links, confidence, time horizon, and watchlist triggers. The agent produces decision-grade output, not exploratory rambling.
Autonomous causal analysis is the infrastructure layer for the next generation of decision support. Abel's agent demonstrates that it's possible: an AI that doesn't just retrieve and summarize, but discovers, validates, and computes cause and effect. The architecture is generalizable — any domain with temporal data and causal structure can benefit. The future of agents is causal.