I feel like the push to agent-first infrastructure will be much less subtle. As computer use gets much better, what incentive does a product team have to duplicate effort across a human stack and agent stack?
Strong piece, the split between "doing" work and "trust/judgement" work is the right axis, and I'm stealing it.
One place the 2030 vision hits a wall that isn't a maturity problem: regulated enterprise. You land on cloud-first as the destination, concurrency, backlogs, refine-from-anywhere, sandboxes that are rock solid. In financial services a lot of that is simply off the table. You often can't send customer data to a third-party sandbox at all; there's data residency, model-in-the-loop audit, and regulators who treat "the agent did it" as your liability, not the vendor's. So agent-first architecture for a bank has to solve the concurrency-and-babysitting problem without the cloud-first escape hatch. That's not a smaller version of your problem, it's a different one.
Which is why "context is the moat" reads even harder from where I sit. In a bank, the context, the semantic layer, what counts as a correct answer, is real advantage and it's locked behind the exact governance that slows the agents down. The moat and the friction are the same wall. Whoever builds 2030-shaped software inside those constraints owns a market your frontier customers never touch.
Our prediction is that there aren't going to be "human-only" users in 2030. People will interact with PostHog through our agents or their agents, we're already a significant percentage of people doing this like 40% of experiments being created through the MCP last week. It's not really us telling them to "go talk to your agents" as it is them preferring to talk to their agent.
I feel like the push to agent-first infrastructure will be much less subtle. As computer use gets much better, what incentive does a product team have to duplicate effort across a human stack and agent stack?
I suspect many teams starting from scratch now will be designing entirely agent-first, yes.
Strong piece, the split between "doing" work and "trust/judgement" work is the right axis, and I'm stealing it.
One place the 2030 vision hits a wall that isn't a maturity problem: regulated enterprise. You land on cloud-first as the destination, concurrency, backlogs, refine-from-anywhere, sandboxes that are rock solid. In financial services a lot of that is simply off the table. You often can't send customer data to a third-party sandbox at all; there's data residency, model-in-the-loop audit, and regulators who treat "the agent did it" as your liability, not the vendor's. So agent-first architecture for a bank has to solve the concurrency-and-babysitting problem without the cloud-first escape hatch. That's not a smaller version of your problem, it's a different one.
Which is why "context is the moat" reads even harder from where I sit. In a bank, the context, the semantic layer, what counts as a correct answer, is real advantage and it's locked behind the exact governance that slows the agents down. The moat and the friction are the same wall. Whoever builds 2030-shaped software inside those constraints owns a market your frontier customers never touch.
but how human-only users would feel in 2030 when companies stop shipping human UIs because "just go talk with your agent?
Our prediction is that there aren't going to be "human-only" users in 2030. People will interact with PostHog through our agents or their agents, we're already a significant percentage of people doing this like 40% of experiments being created through the MCP last week. It's not really us telling them to "go talk to your agents" as it is them preferring to talk to their agent.
Well I won't have a job by 2030. Welp time to go code.