Current trends in AI agent deployment

Software engineering accounted for the top domain in which (Anthropic) agents have been deployed, comprising nearly 50% of total agentic activity, while back-office automation was a distant second at 9.1%. This demonstrates how truly early we are in the current phase of deploying autonomous AI agents.

Deploying agents in production looks very different from deploying them in a test environment. Our own experience building and running agents across investment operations workflows has surfaced a set of lessons specific to the financial domain:

Bias towards determinism. LLMs are most valuable at authoring time — defining logic, structuring rules, generating configurations — but agents in production should execute deterministically wherever possible, with LLM reasoning reserved for genuinely ambiguous decisions.
Agents should be designed around process reengineering, not step automation. Building an agent to handle one stage of a workflow misses the point; the real gains come from rethinking the entire process with cohorts of agents working end-to-end.
Domain knowledge is the difference between an agent that works and one that breaks. LLMs need domain-specific skills to reason correctly on financial data, and deep entity relationships encoded as structured ontologies make a material difference in reasoning quality.
Communication context. Human communicated instructions and intent that govern a workflow must be made explicitly available to agents rather than assumed.
Everyone focuses on agents, but the real differentiator is the orchestration layer that ties workloads together. Agents without disciplined orchestration are a collection of isolated capabilities, not an operational system.

Dreams of autonomous AI agents helping to drive returns

Some of the most advanced thinking around AI applications is currently happening from a front-office point of view. The near-term impact is already visible in how front-office teams work. Tasks that once took months, like turning raw data into novel, predictive features, are being compressed into days or hours, and agentic capabilities are entering the portfolio room as constraint-aware collaborators on investment decisions.

The shift underway is not about deploying AI as a discrete tool. It is about AI becoming part of the operating system for how research and investing work. Make no mistake; it will not be long before agents are deployed to monitor markets, detect non-obvious correlations, optimize portfolio allocations, and assess credit risk. For hedge funds, MCP will be key in helping front-office teams’ proprietary investment tools evolveby enabling conversational access to not only external data sources but also firms’ own proprietary data and operational data.

The implication for any fund building toward front-office AI is that ambition needs to be matched by rigor. Research discipline, auditability, and governance remain as differentiators, regardless of which models a firm has access to.

MCP for lower latency, costs of ownership, and barriers to scale

At this early point in the AI adoption journey, funds leveraging MCP can grab a competitive edge in the AI race by enabling scale of agentic workflows and lowering the burden of integration. For any of these ambitious use cases to drive alpha, the underlying data foundation must be flawless, without which even the most advanced models will underperform or produce inaccurate results.