Corrections fell from almost one in three to under one in ten. Who signs off?
A Dresden founder, two San Francisco co-founders and a Bay Area executive coach have handed the doing to agents, and each draws the line for people in a different place.
In February, almost one in three of a San Francisco co-founder's messages corrected a mistake. By October, fewer than one in ten.
Progress, partly thanks to better models. The person still never left the loop.
A Bay Area coach puts the risk bluntly: agents can report a job done when the record says otherwise.
Four leaders, four setups, one pattern. The doing moves to agents. What stays with the person is where they disagree.
We asked leaders which parts of a job AI has actually taken, and which still need a person. Four answered with specifics: a press pipeline, a coding setup, a quality review and a sign-off habit.
The press hunt is almost fully automated.
Dr. Julia Grosse, Founder of The Cyclical CEO (Greater Dresden Area, Germany)
“In my own company, finding press and podcast opportunities is now almost fully done by AI. An automated task reads every journalist request each day, checks it against my topics, logs the ones that fit and puts the deadlines in my calendar.”
“The better question is what work needs a person, and in which weeks she does it best.”
A clean yes, with no hedging. Then the Dresden founder flips the question: not whether an agent can do the work, but when a person does theirs best. Her rule is simple. Agents take the repetitive work and the planning, and the people's best weeks go to decisions and clients.
Agents write the code. People decide what's worth building.
Isaac Correa, Co-founder of Hellomatik (San Francisco, United States)
“In February almost one in three of my messages corrected a mistake; by October it was fewer than one in ten, partly because the models improved.”
“My messages per hour rose with each new agent up to about seven, then stayed flat, so every agent beyond that simply got less of me.”
Honest numbers, which are rare, and a ceiling nobody puts on the pricing page. The agents write the code. What stays with the person is taste: what deserves to be built, what gets thrown away, when it is good enough to call done.
A successful tool call isn't the same as a happy user.
Raphael Kalandadze, Co-founder & CTO of WANDERO Inc. (San Francisco, United States)
“We've moved the first pass of production-agent quality review to agents. They read conversations, looking for repeated requests and failed tool calls.”
“A successful tool call alone doesn't tell you whether the user got what they asked for.”
Boring, valuable automation: agents do the first read. The catch is evidence. A fix can be technically right and still be the wrong product call, so this CTO wants the task, the evidence and the allowed actions settled before anyone picks who does it.
The agent says done. The record says otherwise.
Eric Rosen, Executive Coach at ericrosen.ai (San Francisco Bay Area)
“An agent will say a job is done when the record shows it isn't.”
“So the doing moved to the agents. The checking and the final yes stayed with me.”
Same pattern, different test. Most debates start with what the tool can do. This coach starts with what a mistake costs. Cheap and easy to undo, let the agent run. Costly or permanent, a person signs off. It fits on a sticky note.
Where they split
Where they agree: the doing moves. Press scouting, code, first-pass quality review.
Where they split is the limit. Dr. Julia Grosse says timing: save decisions and client conversations for the weeks people have the most to give. Isaac Correa says attention: past about seven agents, each extra one got less of the person running them.
Raphael Kalandadze says evidence: a successful tool call alone does not prove the user got what they asked for. Eric Rosen says cost: where a mistake can't be undone, a person signs off.
Four limits: a calendar, a ceiling, a definition of success and the price of a mistake. Same question, four different rulers.
My take
Before choosing a person, an agent or both, I'd ask two things: what does a mistake cost here, and what evidence shows the work is really done? Cheap and reversible, let the agent run. Costly, a person owns the sign-off. That owner is the role worth hiring for.
Your turn
Where does the line sit in your work: timing, attention, evidence or the cost of a mistake? Tell us what you handed to an agent and what you took back.
