By Aditya Jha at Itoflow: Multi-manager hedge funds have an enviable problem. Assets at the platforms tracked by Goldman Sachs increased by more than 25% over the past year and now stand well above $500bn, while measured employee numbers rose by approximately 10–11%. Deployment capacity is now the resulting constraint.
One answer is to recruit more portfolio managers, allocate to external managers, and launch more pods. While all this works, it is labor- and integration-intensive. Every new pod brings people, data, workflows, and controls, all of which must be incorporated into the wider platform. Massive inefficiencies occur when new pods redevelop standardized tools of the trade, such as data pipelines and backtesting and simulation engines.
The multi-manager structure has proven remarkably effective for hedge fund investors. Here, the portfolio manager is central, developing and refining the investment idea with support from analysts and quants. Before deploying the idea as a strategy, it undergoes rigorous validation, testing, and a pilot production run. Largest funds develop the necessary infrastructure to support multiple pods. Central risk functions monitor exposures and enforce limits across pods; some platforms also operate center books that aggregate or offset positions. Part of a fund’s success is building reliable, repeatable workflows and integrations. Still, a new pod may take months to become operational.
Agentic systems offer a route to shortening this build-out. In software development, customer service, electronic design, document analysis and retrieval, scientific research, logistics, and defense, they already coordinate multi-step workflows and increase the output of human teams. Banking and investment companies are beginning to build and adopt agentic systems. For a multi-manager firm, the relevant question is how much of the analytical and operational capability surrounding a portfolio manager can be provided through agentic software.
An AI agent is a software system that works towards an objective. It gathers information and reasons over it. It uses provided tools and can make its own. Unlike a chatbot model, a typical agent can preserve its state across runs, monitor changing conditions, escalate decisions when they fall outside its authority, and absorb user feedback to improve itself. Several specialized agents can work toward a shared objective using common infrastructure, tools, data, and outputs.
From an AI perspective, the pod structure already looks agentic: each pod is semi-autonomous, pursuing a defined strategy within an externally defined risk budget and operating limits. Shared infrastructure provides common analytical and simulation tools, data, and execution. Central functions monitor aggregate exposures and enforce firm-wide constraints.
AI agents therefore offer a plausible mechanism of scalability within a familiar organizational form. They can reduce the replication efforts accompanying each portfolio manager and allow human teams to focus on investment judgment, risk decisions, scenarios, and exceptions.
What can the division of labor look like from an agentic pod’s perspective?
| Humans likely retain | Agents likely carry |
| Ownership of the investment thesis | Research retrieval, hypothesis generation, and evidence assembly |
| Objectives, mandate, and risk appetite | Testing, simulation, and scenario analysis |
| Acceptance of consequential risk | Portfolio-construction and rebalancing proposals |
| Exceptions and changes of policy | Monitoring, authorized actions, and escalation |
| Accountability and investor relationships | Reporting, documentation, and decision history |
Agents can increasingly perform parts of the work currently carried out by analysts, quant researchers, and developers, investigating ideas, writing code, and evaluating results. They can use the firm’s approved data, codebase, and analytical tools while adapting research workflows to a portfolio manager’s strategy and mandate. Agents’ assignments can include:
- Evaluate a signal, checking coverage, timing, and incremental value against a benchmark strategy.
- Probe a strategy hypothesis, testing alternative signals, parameters, or asset universes through out-of-sample and walk-forward evaluation.
- Compare portfolio construction methods and analyze factor exposures, concentration, turnover, and sensitivity to transaction costs.
- Extend or debug research code, add tests, and turn an experiment into a reusable component.
- Monitor conditions and portfolio exposures, triggering re-evaluation when necessary.
- Propose reallocation within the mandate and route exceptions and actions for approval.
One caveat. Morgan Stanley identifies crowding as an essential risk of multi-manager funds. Agents could only boost that risk if different pods still rely on similar ideas, models, and data. But agents could also help with differentiation by exposing non-obvious relationships, plausible second-order effects, and structured creativity, with manager input and within the firm’s mandate.
And at Itoflow, we are building toward this human-agent operating model. The platform already combines quantitative and qualitative research tools with persistent strategy workflows. The Itoflow agent can investigate an investment problem using approved data and supplied or user-provided tools, preserving experiments and decision history, refining and revisiting the strategy through scheduled or event-driven runs. The portfolio manager role is to initiate research, set the objectives, constraints, action permissions, and degree of discretion.
Our direction is to extend that governed continuity across the investment lifecycle: from ideation, research, and validation to portfolio construction, monitoring, and permitted action.
Multi-manager firms will continue to compete for exceptional portfolio managers.
But agentic platforms will increasingly determine how much reliable research, testing, monitoring, and execution capacity each manager can command.
