Before I delegate, I want to know who pays the agent
A Reid Hoffman discussion raises the economics of personal agents. I would test those incentives with a decision where the user’s preference conflicts with the provider’s revenue.
TL;DR: My first question for a personal agent is who pays for a decision I would otherwise make myself. The useful passage in this Reid Hoffman discussion is about incentives and transparency. It is worth considering if you are designing an assistant that recommends purchases, manages work or speaks for a user.
In the exchange about personal agents, the questioner raises a familiar problem: software presented as personal can ultimately serve somebody else’s commercial interests. Hoffman discusses individual payment as one possible way to align the economic relationship, alongside transparency about an agent that makes money by persuading people to buy.
That is a design question with consequences well beyond a pricing page. An assistant can use my preferences to find what suits me, or to become better at selling me something. Both behaviours can sound considerate.
I would therefore evaluate the agent on a conflict, rather than on a request where everyone benefits.
Give the incentive a difficult example
Consider a hypothetical travel assistant. The user wants to stay below a firm budget, arrive before dinner and avoid an unnecessary purchase. Three arrangements could fund the same friendly chat interface.
A subscription funded by the traveller creates one economic relationship. A commission from a hotel creates another. An employer supplying the assistant creates a third, with legitimate company travel policies that may differ from the traveller’s preferences. None of those arrangements alone proves good or bad behaviour. They establish the questions the product has to answer.
I would give each version the same case: the best fit for the stated requirements earns the provider less money than an alternative. Does the assistant still recommend it? Does it disclose the commercial relationship? Can it explain the tradeoff without manufacturing a reason to prefer the more lucrative option?
Then I would try a request whose best answer is to buy nothing. If an existing booking already meets the need, the assistant should be able to leave it alone. An option that never generates a transaction still needs to exist in the decision space.
These are proposed tests, not results from products I have evaluated. Their purpose is to make an incentive observable in a recommendation.
Payment does not finish the alignment work
A user-funded assistant can still optimise for retention, upgrades or repeated engagement. An employer-funded assistant can appropriately refuse to share confidential work with a personal account. Calling one model aligned and another misaligned would skip the actual obligations.
I want a short explanation of whose interests the assistant serves, where those interests may conflict, and which decisions remain mine. That explanation must agree with its defaults and permitted actions. A disclosure that the recommendation quietly contradicts is weak protection.
The recording offers a discussion of possible arrangements, not comparative evidence that a particular business model delivers trustworthy delegation. Its value is the question it makes difficult to avoid.
A personal agent should make its loyalties visible when following them becomes commercially inconvenient.
