Balyasny built the evaluation pipeline and data governance before deploying across roughly 95% of the firm’s 180 investment teams. If you’re evaluating your own readiness to deploy agents at scale, the question to ask is whether your data foundation would hold up under the same scrutiny.

Build for autonomy in stages

The most expensive mistakes happen when you give agents too much autonomy. An agent that recommends a position adjustment carries different risk than an agent that executes one with settlement finality and client-visible outcomes.

Start by letting the agent observe. It watches the workflow, flags what it sees, and you evaluate whether the flags are useful. Once you trust the observation, let the agent recommend a course of action. You still make the call.

After enough accurate recommendations, you can let the agent act with your approval before execution. Full automation, where the agent acts without a human checkpoint, only makes sense for workflows where the cost of a wrong action is low, and the agent has a long track record in the earlier stages.

Consider a portfolio manager rebalancing a multi-asset portfolio. At the first stage of the workflow, the agent could monitor drift from target allocations and flag when a portfolio exceeds tolerance. The agent’s output leaves the decision on how to proceed with the PM.

At the next stage, the agent would propose a rebalance with specific trades, sized and sequenced, accounting for tax lots, transaction costs, and restricted securities. The PM reviews the proposed trades and approves or modifies before routing the orders for execution.

In a fully autonomous workflow, you might let the agent execute within predefined parameters, such as no single trade above a set notional and only during normal market hours, with a PM reviewing after the fact. An audit trail would give compliance officers, client reporting teams, and operations staff a complete record of what the agent did and why. This represents the next stage of AI evolution, but we’re not seeing firms use it in production yet.

Each stage of proven reliability gives you confidence to move the agentic component downstream in the workflow. With that record in place, a promising pilot becomes a permanent part of your infrastructure.

From prototype to production

Given the speed of agentic AI deployment and innovation, trying to follow the technology feature by feature is a lost cause. A principles-based approach lets you avoid confusion, paralysis, and drift. Applying all four of these principles to every initiative gives each agent the best chance of reaching production.