The Roman hierarchy Botha and Dorsey describe existed because a leader can hold only a few people in mind. Agents remove that constraint for machines. They do not remove it for us. If we pretend otherwise, we rebuild middle management as notification hell.
Last year everyone said “agents.” Few could say what they were without pointing at a demo that fell over.
This year the shape is clearer. An agent is not a chatbot with plugins. It is a persistent worker with memory, tools, and a job. OpenClaw showed the frontier. Hermes made it livable for people who liked technology but did not want to live inside a terminal. Then Grok Bot arrived, and I needed a few days to notice it was not another clone.
The product that wins is not the cleverest weekend project. It is the lab that ships agents as a product: cloud-native, personal, and polite enough that a human can stay in the room.
From chewing glass to a product
Peter Steinberger’s OpenClaw was the pioneer moment. It proved that a personal agent could feel like a companion rather than a form. It also chewed glass. Installations, permissions, broken loops, the kind of pain that only heavy developers treat as sport.
Anthropic then made a choice I still think was a mistake for the agent frontier, even if it made sense on a spreadsheet. It pushed agent power users out of the comfortable subscription lane. Compute was scarce. Enterprise paid. The scouts who were wiring models into real life went shopping.
Steinberger joined OpenAI to bring agents to more people. OpenClaw stayed open source under a foundation. Independence on paper. In practice, for many of us, the glass remained.
Hermes won the next round for a simpler reason. OpenAI OAuth, then Grok OAuth, and a harness that improved faster than the model gap. Models still mattered. They mattered less as the operating system around them got better.
I wrote that shift in When Models Converge, the Factory Wins. Grok 4.1 and 4.2 were not yet good enough for the full stack. GPT-5.5, and especially GPT-5.6, stole the daily show. With Grok 4.6 the differences keep shrinking. Intelligence becomes modular. Price, speed, access, and product design become the fight.
Then came Grok Bot.
Not another local copy
It took me a few days to understand why it felt different.
It is not OpenClaw with a nicer coat. It is built on Cursor’s cloud-agent architecture. Every agent is native in the cloud from the first minute. No install ritual. No desktop as the only stage. No “my laptop went to sleep and the company stopped.”
Two other differences landed harder than the architecture diagram.
The first is personal UI. Not a generic chat pane pretending to be a person. Named agents. A place that feels like yours.
The second is human communication from the start. Short. Clear. No verbose tech soup. No allow once / always questions about something I do not understand and should not have to. OpenClaw was for the heavy tech developer. Hermes was for the tech enthusiast. Grok Bot is the first one my kids can also follow.
That is not a soft feature. That is distribution.
Adoption showed it. Marketplaces. Templates. In my own hands: smart handling of secrets, inventive problem-solving, and speed. In a few hours it turned my essay site from a paid service into a cleaner site shaped for me. That is the real deflation. Not a cheaper plan. One less tool I pay for.
Because it is personal, sharing secrets feels less insane. Mail and other accounts can be decoupled into agents that hold the keys for a purpose, not into a random bot in a public channel.
I still struggled with one design choice.
The individual and the swarm
Grok Bot is built around individuals. My Hermes agents were opened through Slack and Telegram toward other agents and other humans. That swarm felt closer to how a company actually moves.
I had also made a thinking error.
I believed humans and agents should meet as equals on one platform. Roelof Botha and Jack Dorsey’s essay From Hierarchy to Intelligence still has my vote on the deep point: AI should change how we coordinate, not only how fast we type. Block’s Buzz tries to make agents first-class citizens next to humans. Telegram already treats bots more natively. In Slack, agents remain second-class seats with a costume.
I still want agents to have full membership where work happens.
The error was elsewhere.
Agents produce far more than humans can absorb. Equality of identity is not equality of throughput. In our company, agents can clear something like 1,500 issues in three days while humans do spot checks. Put that flood into a human chat product designed for human gossip and you do not get collaboration. You get a firehose aimed at a face.
I also spoke with Ando in San Francisco, another Slack-killer bet for agent-pilled teams. The category is waking up. The hard problem is not “can the agent post.” The hard problem is “can the human still think.”
Parallelism kills the desktop myth
There is a second move happening at the same time.
Because we can run many processes in parallel, the desktop becomes the wrong factory floor. Your CPU and GPU are bottlenecks. Sleep mode is a hard stop. Cursor’s cloud agents keep going. They scale with tokens, not with whether your lid is open.
So two human jobs expire together.
We have to get off orchestration as a daily identity. And we have to define, with uncomfortable honesty, when a human still adds value.
I already wrote the answer once in The Last Human in the Loop. Humans give meaning to what agents do: truth, curiosity, beauty. Last week we rebuilt our system so it can develop more autonomously, and so humans get clear lists for human work.
One list is substance: decisions, usually as A, B, or C. Or a stuck process that needs a key, a login, a judgment no model should fake.
Another list is eyes: tests and human review where reality still needs a body in the room.
That is not a platform of humans and agents sharing equal airtime. It is closer to WhatsApp. Occasionally I share a message with a colleague I actually read and act on. Agents and bots must know that human attention is scarce. Channels stay few and logical. Agent messages stay short, clear, and explicit about what is asked.
When attention is thin, a good agent holds back. When focus returns, it accelerates. The agent becomes sensitive to humanity, so the flow of information can still travel through the channels humans already use.
The Asana warning
Years ago I introduced a CEO to Asana. He was delighted. Finally a place to empty his head. He kept adding and structuring tasks. After a few weeks he came back frustrated. Nobody picked them up. Why weren’t people working?
They were already busy. Thirty new tasks with your name on them do not create flow. They create apathy. The board fills. The body freezes.
Agent platforms will make the same mistake at machine speed if they treat human inboxes as infinite.
The Roman hierarchy Botha and Dorsey describe existed because a leader can hold only a few people in mind. Agents remove that constraint for machines. They do not remove it for us. If we pretend otherwise, we rebuild middle management as notification hell.
What I am actually watching
So the scoreboard, for me, is no longer “who has the smartest model this week.” I watch who ships agents as a product ordinary people can live with, who runs them in the cloud so parallelism is real, and who treats secrets and personal context as sacred rather than as a dump. I watch who gives agents first-class tools without forcing humans to drink the output hose, and who designs for attention: hold when we are gone, move when we are present, ask only when a human decision is the work.
OpenClaw opened the door. Hermes made a home for builders. Grok Bot is the first version that feels like it might leave the pioneer class and enter family and company life without asking everyone to become a sysadmin.
The next fight is not agent versus human. It is whether we keep humans as orchestrators of noise, or as the scarce source of meaning inside a swarm that can finally outrun us.
That is the line.