Autonomy is earned.

Giving a single AI model the keys to a fund would be a bad design. A model that reads the market, decides and signs its own trades has nothing to catch its mistakes, and its manager could not explain its decisions to an investor afterwards.

In the system we are building, the agent reads and proposes around the clock, across more positions than a team could watch. Whether a proposal becomes a trade is decided outside the agent, by the mandate, the checks below and the people who supervise it.

The mandate comes first

Every agent works under one written, versioned mandate that sets its strategy, its risk limits and how much autonomy it has. People write the mandate, and only people can change it.

Inside those limits the agent has room to judge. It chooses when to rebalance, which yield band to aim for and how quickly to unwind a position that is getting worse. If a lending market’s utilisation keeps climbing, it can start easing out before a hard limit forces the exit. Approving a new market, raising a limit, changing fees and governance all stay with people.

Every decision is a proposal

The agent reads live state: positions, liquidity, rates, collateral and the risk signals that move them. From that it proposes to add, reduce or move capital, or to hold. Each proposal comes with a written reason and the evidence behind it.

A decision to hold is recorded like any other. When data is stale or missing, the agent lists what it could not check and holds.

Four checks before anything moves

Each proposal passes four checks in turn before any capital moves. Any one of them can stop it.

An agent can argue for a trade. It has no way to argue a limit out of the contract that enforces it.

Keys never reach the model

Private keys never enter a model’s context. Each agent has its own identity and credentials and reaches only the data and tools its role needs. A change to what an agent may do is a reviewed change that a person approves, so every permission has an author and a date.

Autonomy is a ladder

Every agent starts at the bottom, in shadow mode, and moves up one level at a time.

Recommend. The agent runs beside a rules-based baseline, on live data and on replays of hard days: a large redemption, a sudden shift in yield, stale data, an illiquid market, a depeg. It executes nothing, and every proposal it would have made is logged.

Human-approved. Proposals that pass the guardrails and the reviewer go to a person, one trade at a time, on a small amount of capital.

Bounded execution. Trades inside a tighter limit run on their own. Larger or unusual trades go to a person.

To move up a level, an agent needs the team’s sign-off against evidence. It must have executed no mandate breach, and matched or beaten the baseline’s return with liquidity at least as good. Its log must explain every decision on its own. And it must have handled at least one case better than the baseline did, such as easing out of a market before a limit forced it.

Every decision leaves a record

Each proposal is logged with the mandate version, the inputs the agent saw, the alternatives it weighed, the checks it passed or failed, who approved it and the outcome. Blocked, rejected and hold decisions go in the same log. Our standard is that anyone responsible for the fund can explain any decision from the log alone.

One agent, then a fleet

A second strategy gets its own mandate and its own agent, and runs through the same four checks and the same ladder. The portfolio manager’s work moves to writing mandates, allocating capital between strategies and deciding when each agent is ready for the next level.