Financial agents need data platforms and control systems, not better models.

The core argument. AI engineering in financial services has moved past chatbots into production agents, but the real work is organizational, not model-based. The conference's three keynotes—from Bridgewater, Anthropic, and Two Sigma—converge on the same hard truth: models are the easy part. The hard part is building data platforms, agent foundations, and control systems that work at scale in regulated environments. Each talk offers a distinct piece of this puzzle, and together they form a usable playbook for engineering leaders.

Bridgewater's lesson. Bridgewater's Vault platform ingests over 100 million documents from 135 external sources plus internal memos. The team learned that humans and agents have fundamentally different needs from a data platform: humans need high precision (few relevant documents), agents need massive recall and scale. Their solution was TurboPuffer's hybrid filtering, which gave them both speed and recall. The real prize, though, is making agents proactive rather than reactive—the Investor Workbench keeps agents always on, tracking what matters and pulling in humans only for decisions.

Anthropic's framework. The Anthropic talk proposed a five-layer foundation for scaling agents: connectors (via MCP), skills (versioned business rules), identity/permissions, approval/sandboxing, and evals. The key insight is that models improve every four months while adoption lags—creating a "diffusion gap" that firms can exploit. Their Claude for Financial Advisors product demonstrates the pattern: 25 pre-built connectors, eight skills on GitHub, and a trust layer where the advisor approves everything.

Two Sigma's control system. The Revlon $900 million payment error is a cautionary tale for agent autonomy. Two Sigma's approach is a five-part control system: an agent registry (visibility), risk-based scoring (correct review intensity), intent-declared tool access (granularity), shared evals (repeatability), and runtime classification (sustainable human-in-the-loop). The critical insight is that safety and innovation are not a tradeoff—controls are brakes that let you go fast safely. The design principle is that agents should declare their purpose upfront, and policies should be enforceable at runtime.

The disposable software thesis. Capital One's Max Kanat-Alexander introduced a framework distinguishing durable software (years of maintenance, many dependencies) from disposable software (hours to rebuild, few users). Code quality matters only for durable systems. The new discipline is making software more disposable through isolated environments, enforced independence, and redundancy. For engineering leaders, the implication is that not every agent needs enterprise-grade guardrails—some are calculators for one person, not trading systems.