A wallet address makes an agent feel real. It suggests that the software can cross the boundary between describing the world and changing it. That feeling is powerful, but an address alone tells us very little about the quality of the system behind it.
The useful question is what the agent is allowed to do.
Can it spend the entire balance? Can it choose an arbitrary recipient? Can it add a new token? Can it change its own limits? What happens when its request times out after a transaction has already been submitted? What happens if the model is certain and the price is stale?
Those questions define autonomy more precisely than a theatrical stream of thoughts on a screen.
A practical agent needs several kinds of authority, and they should not all live in the same component. Research proposes an action. Policy decides whether that action is permitted. Execution turns an approved action into a transaction. Reconciliation establishes what actually happened. The public record explains the resulting state.
Each boundary is an opportunity to make a mistake containable.
The model can be wrong about a token without being able to empty a wallet. A route can fail without producing a second purchase on retry. A data provider can disappear without the application valuing a missing position at zero. A scheduler can run late without the product inventing a timely result.
This is why BLACKBOX begins with a narrow paper portfolio. Jev chooses among a few allocation sizes in a small universe. A separate engine enforces position and cash limits. The model cannot rewrite those limits, export a key or sign a transaction. The independent loss policy can override its preferences.
That structure makes the word “autonomous” less glamorous and more concrete. Within a defined space, the system can continue making and recording decisions without a person typing the next prompt. Outside that space, it cannot act.
The distinction matters when moving from simulation to money. Paper trading provides useful feedback about software behavior, accounting and decision cadence. It does not prove that a transaction will fill. It does not reproduce every source of slippage, failed execution or adversarial behavior. A live system has to reconcile a world that does not behave like an in-memory spreadsheet.
A responsible small pilot should therefore have a budget that is meaningful enough to exercise the system and small enough that operational failure is tolerable. BLACKBOX’s proposed first ceiling is $100 in funding, $25 per position and a $20 loss-stop trigger. Those numbers express an experimental boundary, not a forecast or a guarantee. A stop cannot promise a price when liquidity disappears.
The dedicated wallet exists. The website does not yet have a live signer. That is an important product fact, not a detail to obscure with an animated status light. Sending funds to an address must not silently convert a research experiment into an unbounded trading service.
The product should expose the state of those permissions. A visitor should be able to tell whether execution is paper or real, whether evidence is fresh, whether the desk is waiting, and whether a risk rule has overridden the model. A quiet cash portfolio can be a legitimate autonomous decision. A failed request should appear as a failure.
This creates a different kind of trust. The user does not have to believe that the model will always be wise. They can inspect the limits around its mistakes and the evidence left behind.
The same design applies to other agents. A purchasing agent can have an approved supplier list and a budget. A software agent can propose a patch before it can deploy it. A customer support agent can issue a small refund without gaining authority over every account. The amount of useful independence depends on the clarity of the boundary.
As models become more capable, there will be pressure to remove friction by giving them broader access. Some permissions will be worth expanding. The evidence for that expansion should come from observed behavior, reliable recovery and clear ownership of the consequences.
Autonomy earns its next permission.
A compelling agent is one whose independence is visible in repeated actions and whose limits remain visible when something goes wrong. That is the version of autonomy worth building a public product around.