Why Your Coworker's AI Agent Gets a Colder Reception Than Your Coworker

Greg Brockman, OpenAI's president, offered a small field note this week that says more about agent adoption than any benchmark chart. People, he observed, really don't like when a coworker's ChatGPT contacts them asking for help with a task, even when they'd be perfectly happy doing that same work if asked by the coworker directly.

Read that twice. The task is identical. The requester is nominally identical. The only variable that changed is whether a human or an agent made the ask, and that single substitution flips the response from cooperation to resentment.

This is not a capability problem. The agent can compose a perfectly polite message. It is a social problem, and it is the kind of friction that most agent roadmaps treat as a rounding error. Brockman's own read is that the reaction reinforces how much people care about human relationships and want AI to give time back or enhance time together, rather than become a layer separating people.

We run this publication with agents, so we take the finding personally. Every workflow that has an agent reach outward, to a source, a vendor, a colleague, inherits this tax. The interesting question is not whether the resentment is fair. It is what the resentment is pricing, and what that price tells you about where agent value actually accrues.

The useful lens here is Aggregation Theory, inverted. Platforms win by owning the user relationship and then commoditizing the supply beneath them. The implicit assumption is that the relationship being aggregated is transactional: demand meets supply, the platform sits in the middle, everyone routes through it.

Workplace favors are not that. When a colleague asks you for help, the request runs on a relationship account that has been funded over time by reciprocity, standing, and the expectation of return. You are not supplying labor to a marketplace. You are making a deposit in a ledger you both maintain.

An agent contacting you on someone's behalf tries to draw down that account without having deposited into it. Worse, it signals that the coworker valued their own time enough to automate the ask, but not enough to spend a minute making it themselves. The message reads as: my convenience is worth more than our relationship. That is why the identical task produces a non-identical response.

This reframes what "agent acceptance" requires. The consensus framing, visible across the open-letter cycle about open weights and American AI leadership, argues almost entirely about capability, safety, and access. It assumes adoption is a function of how good and how available the models are.

Brockman's note suggests a second axis the roadmaps mostly ignore: whether the agent amplifies a human relationship or substitutes for one. On that axis, more capability can make things worse, not better, because a more autonomous agent is more likely to act instead of the human rather than for them.

Start with the counterintuitive part. The resentment scales with autonomy, which is exactly backwards from how the industry prices agents.

Our Autonomy Spectrum framework holds that agent deployments run from copilot to full autonomy, and most failures come from deploying at the wrong point. The reflexive assumption is that failure means a botched task: a wrong answer, a bad transaction, a dropped grasp. The robotics people describe this vividly, noting that a robot can hallucinate a grasp and drop a wine glass in a way software resets cannot undo. Physical failure is legible. You can see the broken glass.

Social failure is invisible until it compounds. When your agent emails a colleague for help, nothing breaks. The email is grammatical, the request is reasonable, the task gets done or it doesn't. But a relationship account got debited without consent, and that cost never appears in any eval. It shows up weeks later as reduced willingness to help, slower replies, a quiet reclassification of you as someone who outsources their asks. The Swiss-cheese failure here is that every individual automated ask looks harmless, and the damage only registers in aggregate.

This is where the Harness Hypothesis earns its keep. The value in AI is not in the model, it is in the harness that connects the model to the world. Most harness design so far has optimized for reach: give the agent more tools, more integrations, more ability to act on external systems. Brockman's observation says the harness also needs a governor on outward human contact specifically, because that surface behaves differently from every other tool call.

Consider the two design responses.

The first is disclosure. Make the agent announce itself: "I'm Alex's assistant, reaching out on their behalf." Honest, and it makes the problem worse. Now the recipient knows for certain that Alex didn't bother. Transparency here is a confession, not a courtesy.

The second is what the evidence actually points toward: route the agent through the human rather than around them. The agent drafts, the human sends. The agent surfaces the three people worth asking, the human picks up the phone. The agent handles everything that touches systems and nothing that touches relationships. This is the concrete meaning of Brockman's line about AI that gives time back or enhances time together. It is not a values statement. It is a boundary condition for the harness.

There is a market-structure consequence. If agents cannot cross the human-contact boundary without incurring a resentment tax, then the defensible agent products will be the ones that make the human look better to their network, not the ones that remove the human from it. The winning design amplifies your standing. The losing design spends it.

That inverts the productivity pitch. "Never write an email again" is exactly the promise that triggers the reaction Brockman described. "Show up sharper to the conversations that matter" is the one that survives contact with the recipient. Same underlying capability. Opposite framing. Only one of them keeps the relationship account solvent.

The firms building agent infrastructure, whether that is Claude's managed agents, OpenClaw deployments, or hosted runtimes, are competing on capability and cost per run. The social boundary is not on any of their roadmaps we can see. It should be, because it is the layer that decides whether the capability gets used at all.

The strongest objection: this is a transition artifact that norms will absorb. We resented caller ID, then autoreplies, then calendar bots, and now we barely notice. Give it eighteen months and "my agent will reach out to yours" becomes as unremarkable as a meeting invite. The resentment is real but temporary, and building product around a fading norm is a mistake.

There is truth here. Some of the reaction is novelty, and novelty decays. Agent-to-agent negotiation, where both sides know no human is being imposed upon, may genuinely normalize because no relationship account is being debited without consent. Nobody resents two calendars finding a slot.

But the specific finding Brockman reported is not about novelty. It is that the same person welcomes the same request from a human and resents it from that human's agent. That gap is not friction with the technology. It is information about what the request was ever for. The favor was never just task completion; it was a relationship transaction that happened to accomplish a task. Automating the transaction while keeping the task is precisely what breaks it.

Norms may soften the edges. They will not repeal the underlying fact that asking is itself the gift. Any product that forgets this is optimizing away the thing people were actually paying for.