For financial technology institutions operating at enterprise scale, the initial wave of Generative AI adoption, which was centered on isolated prompt boxes and greenfield code generation, quickly hit a ceiling. Within Bloomberg’s 45-year-old technology estate, where more than 9,000 software engineers manage massive JavaScript codebases, global financial data streams, and thousands of Terminal applications, simply producing code faster is not enough.
According to Andrey Rybka, Head of Architecture and Emerging Technologies (and AI Security) in the Office of the CTO at Bloomberg, moving from simple automation to enterprise transformation requires stepping into a new operational paradigm: Multiplayer AI.
The Four-Stage AI Continuum
Drawing from his 13 years at Bloomberg and three decades in the software industry, Rybka frames enterprise AI adoption across four distinct evolutionary stages:
- ChatGPT-Style Prompts: Single-input conversational queries and baseline assistant interactions.
- Connectors and MCPs: Standardized integrations through protocols like Model Context Protocol (MCP), linking models directly to telemetry, metrics, and internal toolsets.
- Codified Agentic Harnesses: Systematized agent loops that automate repetitive workflows and execute multi-step operational tasks autonomously.
- Multiplayer AI: Shared contextual environments where human engineers and autonomous agents actively collaborate on strategy, specifications, and execution loops.
“The majority of our work is a loop,” Rybka explains. While single-prompt boxes require constant human triggering, the ultimate destination is an environment where humans set directional outcomes and agents collaborate within shared context to execute and verify complex workflows.
Multi-Model Arbitrage and Agent Hierarchies
Building effective agentic systems inside a regulated financial services environment requires rethinking model selection. Rather than relying on a single monolithic frontier model, Rybka advocates for structured agent hierarchies and multi-model arbitrage.
In his own personal systems, Rybka orchestrates a hierarchical structure overseen by a “Chief of Staff” supervisor agent, complemented by specialized adversarial agents that continuously audit and check worker outputs. By benchmarking frontier and open-weight models across distinct task phases, such as planning, execution, and adversarial testing, architects can match the optimal model to each phase. This multi-model approach yields significant performance gains while keeping operational costs tightly governed as a core design constraint.
“I write a spec or co-create the spec with AI naturally as a co-collaborator. The best agent I was able to create used multiple models and achieved really good cost efficiency.”
— Andrey Rybka, Head of CTO Architecture & Emerging Tech, Bloomberg
Shifting the Cost Function and Expanding Who Builds
This shift in human-agent collaboration is fundamentally reshaping software economics. As Lei Zhang, Head of Technology Infrastructure Engineering at Bloomberg, highlights, generative tooling fundamentally alters the “cost function of software engineering” — changing the calculus of which technical projects and maintenance efforts are worth pursuing.
One of the most striking outcomes at Bloomberg is the demographic of new software creators: non-technical employees. Armed with baseline specification templates, custom LLM skills, and structured toolsets provided by architecture leaders, non-technical staff across the firm are actively building custom production automations for their daily workflow loops.
“If you give them the baseline toolset, people are just so creative and can produce amazing automations,” Rybka observes.
Grounding, Guardrails, and Core Fundamentals
In the finance sector, speed cannot come at the expense of reliability or compliance. To prevent hallucinations in financial applications where “fraud is fraud, with or without AI,” Bloomberg enforces strict factual grounding. Agentic systems first query trusted, curated corpora using vector search and Retrieval-Augmented Generation (RAG) before applying LLM reasoning, ensuring every output remains traceable to verified facts.
Furthermore, Rybka emphasizes that agentic deployments must embed comprehensive defense-in-depth upfront — including sandbox execution environments, ingress/egress controls, tool and action governance, identity and trust frameworks, prompt defense, data protection, and observability.
For AI engineers entering this new landscape, Rybka’s advice returns to core software engineering principles:
- Master the Fundamentals: Do not skip computer science basics — rigorous requirements gathering, system architecture, and testability are more critical than ever.
- Design for Non-Functional Requirements: Treat security, performance, scalability, reliability, and observability as core architectural primitives rather than late-stage ticket add-ons.
- Focus on Intent over Implementation: As AI takes over step-by-step code writing, the engineer’s value shifts toward specifying what must be achieved and establishing deterministic verification.
The Road Ahead at AIE NYC
As financial institutions modernize, the lesson from Bloomberg is clear: AI is no longer just a passive coding autocomplete — it is a collaborative partner. By combining multi-model agent hierarchies, strict data grounding, and robust security harnesses, AI engineers can safely scale autonomous agents from simple scripts to full multiplayer co-creation.
Top comments (0)