@bot is the iPhone moment for agents.
The form factor is the magical part. It's like iMessaging an agent that has context to tools you use and learns over time, all in the cloud.
Models will change, but this form factor is here to stay.
Slow in the corners. Fast on the straights.
In F1, the fastest drivers take the right line slower than it feels, because that carries speed onto the straight.
Same with AI. I could answer most emails in two words, fast in the corner, but the context evaporates and the next time I start from zero.
Run it through the AI system or bot instead. A few extra minutes, slow in the corner, but the memory and context compound. That's the line that builds in the straight.
Most people optimize the corner. The leverage is in the straight.
@elonmusk “The most common mistake of smart engineers is to optimize a thing that should not exist."
With AI we can automate anything in an afternoon, but that does not mean we should.
The skill right now is not using AI on everything. It’s knowing what to leave alone.
@ChristianXCesar@bot Approvals can be at both, but I’m trying to turn my skills into Task bots and have the approval for auto execution there. I build enough trust with a Task bot doing one flavor of workflow that I trust them to auto send the email follow up for me.
Don't Manage the Agent. Manage the Diff.
An agent is a bundle of jobs and we need a human in the loop until trust is earned.
I'm using Grok Bot (@bot) with three seats: Chief of Staff (routes; judgment stays), Employee bots (judgment, memory, skills), Task bots (narrow job). Trust links to the job, not the seat.
Measure the Diff
The Diff is the distance between the draft and what was shipped. I don't manage the agent. I manage the Diff. But the Diff can lie. Zero Diff may mean more trust, or that I stopped reading. It measures agreement, not correctness.
Ask: Did I agree with the reasoning? Did the outcome work? What would break if wrong? The edit Diff shows whether it sounds like me. The outcome shows whether it works.
Scoped Autonomy
When Diff and outcomes are aligned, pull the approval gate for that job. The bot acts, then reports; I review the recap. Autonomy must be earned, scoped, and reversible; the bar scales with blast radius. My Chief of Staff doesn't grant autonomy to a seat. The Diff grants it to a job.
The Point
Measure, shrink, prove, then remove the human in the loop only where blast radius allows and the Diff shrinks to zero. That's how a @bot earns trust.
@emaramigin@bot@calcom@attio Good question, for now, it just returns an update after the deal is created to Scout (my sales agent). I don’t have that “New Deal” task bot do anything else at the moment other than listen for the @calcom webhook and execute the workflow.
Two ways people are structuring @bot agents: employees and task bots, with a chief of staff at the top to do the routing and delegation.
Employees own a domain with judgment, memory, and skills. Think a customer support agent. Task bots own a narrow job with a thin persona but thick, repeatable skills. Think an email inbox triage agent.
Employees help steward a function of the business. Task bots help by redeeming time on task clutter.
If a task bot wants a mission, promote it or kill it. If an employee collapses into one workflow, demote it to a task bot.
Did the next step!
Took one Employee skill (@calcom demo booked → create/enrich @attio deal) and created a Task bot for it. Same skill file, now with a thin, specific @bot persona.
"Scout" (the Employee agent) still owns judgment after the record exists. The Task bot redeems the clutter of create/enrich so Scout doesn’t wake on every webhook.
Seems like this is the path I'll keep exploring: turning Skills into Task bots.
Expanding on this idea...
I think the evolution of skills (readable SOP's) are creating specific task @bot 's to execute those skills. Same skill file, plus memory, connectors, and context around that one workflow. Narrow use case, but better execution.
Chief of Staff delegates to Employees with a domain and Task bots run the workflows.
Overkill? Maybe, but that's what I'm testing right now.
Next step for me: take an Employee skill and turn it into a specific Task bot.
Most AI usage is Productivity Theater.
Lean AI is a framework that filters every AI use case through one question: does this impact revenue?
If it doesn't, cut it.
https://t.co/hMR4cTb48k