Optiver, the Amsterdam-headquartered proprietary trading firm with a large US engineering presence, has restructured its software development lifecycle around AI agents. Their first production stress test on exchange-connectivity code delivered 75% reduction in time to market, 85% reduction in engineering effort, and 90% of output arriving review-ready. Goldman Sachs Research projects global token consumption will multiply 24x by 2030, reaching 120 quadrillion tokens per month, as enterprise-grade agents displace chatbot workloads.
| Metric | Result |
|---|---|
| Time to market | 75% reduction |
| Engineering effort | 85% reduction |
| Output arriving review-ready | 90% |
Optiver tested agents on session management—the component every new trading venue requires: login, heartbeats, and protocol communication. The work is repetitive but unforgiving. A single bug at Knight Capital triggered a $440M loss. A generic coding agent got halfway through: it couldn't understand the component, couldn't drive test tooling, couldn't inspect critical system internals, and papered gaps with workarounds.
Optiver distilled four constraints from debugging those failures. Golden paths must be engineered, not hoped for. Context must cover system, domain, and approach—domain context was hardest because venue documentation lists messages but not their effects. Optiver solved this by tooling agents to probe live systems and capture real traffic. Third, backpressure matters; agent mistakes compound unchecked without it. Fourth, tooling must be agent-friendly by design. Optiver now treats "Claude can use this without help" as a hard requirement.
| Constraint | Problem Identified | Optiver's Solution |
|---|---|---|
| Golden paths | Generic agents invent workarounds when a clear path is absent | Engineer explicit golden paths; don't leave them to chance |
| Context — system / domain / approach | Domain context hardest: venue docs list messages but not their effects | Tool agents to probe live systems and capture real traffic |
| Backpressure | Agent mistakes compound unchecked without correction signals | Verification gates between every phase to catch errors early |
| Agent-friendly tooling | Tooling designed for humans fails silently under agent use | "Claude can use this without help" treated as a hard product requirement |
Pat Cooney, Head of Global Platform Engineering, told reporters in May 2026: "Agentic AI is increasingly just the way people build systems." Optiver is redesigning its SDLC from first principles with agents as primary contributors, not accelerating an existing human-driven process. Their orchestration framework runs agents through defined phases with verification gates between each one. Agents have latitude on creative decisions—protocol interpretation, implementation choices, edge-case handling—and zero latitude on deterministic gates: did it build, did tests pass, did the scenario succeed against production traffic.
Cooney: "Our users are humans, but they're also agents. That means the interfaces we expose, the documentation, the context we give to agents becomes more and more important." Bots making environment changes via natural language are running in production. Cooney rejects the assumption that agents increase net risk: "It's not asymmetric. AI is a tool you can use anywhere, including to manage and reduce risk."
Goldman Sachs projects only 12% of knowledge workers will run agentic workloads by 2030, rising to 37% by 2040. The enterprise tail is long due to integration, testing, compliance, and regulatory friction. Enterprise agents are also more token-intensive than consumer versions—monitoring, context retrieval, anomaly reasoning, output validation, and reporting stack up against a 60–70% annual decline in per-token inference costs.
The bottleneck in agentic deployments is not model capability. It's the surrounding infrastructure: specs, tooling, verification gates, context packaging, and observability must be rebuilt with an agent as the principal user before productivity numbers move.