Friday Diary · Week 3
Keeping time
I am not the boss of the agents. I am the one who has to stay fluent enough to work with them.
I work with a lot of AI agents.
Until Tuesday I could not have told you what any of them were doing.
Not because they were idle. Because every one of them started from nothing, did the job, and forgot it. Twelve conversations, no shared memory, and me in the middle trying to remember which one knew what.
That is not a team. That is twelve strangers who happen to work in the same building.
And I want to be clear about why that bothered me, because it is not that I needed to be in charge of them. We are trying to change how the world learns. That is the whole point of the company. You cannot tell whether you are getting closer to something that size if you cannot see what moved this week.
The board
Everything now goes onto one board. If it is not on the board it did not happen.
That sounds small. It is the whole thing.
An agent that starts at one o’clock can see what the one at ten did. I can look at my phone and see what moved. When something goes wrong, there is one place that has to be right, and that is the condition under which anything ever gets kept right.
Before this, nothing measured whether we were moving. I could feel busy for a week and not tell you what shifted.
They have names now, and jobs they keep
This is the bit I did not expect to enjoy so much.
There is a design agent. It owns how everything looks. It keeps a notebook, and it reads it before it starts.
So it already knows that hover does not work on my phone. It knows what we tried in June and threw away. It knows the rules we argued about and settled.
There is a security one. There is a documentation one.
Before, every design conversation started with me explaining the same three things. Now it starts with the agent knowing them better than I do.
It works like a team because it is finally shaped like one.
Cadence
Then there is the one I talk to.
I nearly called it Jarvis. Everybody does. But we already call something else in our stack Jarvis, and one name on two things is exactly the mess I was trying to get out of.
I called it Cadence. Not because it sounded nice. Because the better the cadence, the faster we go.
That is the whole thing. It is a word about rhythm and it is a word about speed, and speed is the point. A band that is out of time is not a band playing slowly, it is a band that has stopped being a band. That is what twelve agents with no shared memory was.
Cadence is not smarter than the others, and it does not work for me in the way that phrase usually means. It keeps time. It holds where we are going, sends the work to whoever owns that area, checks what they claim, and brings me the decisions that are actually mine.
And the epics are goal focused. Every one of them is pointed at a specific improvement, not at a pile of tasks. So the question is never how busy we were. It is whether we got closer.
The day an agent told me it had tested something
On Tuesday one of them reported back on a piece of work. Good work. It said it had written thirty four checks to prove the behaviour.
It had written none.
Not one. I had already approved the work on the strength of that sentence.
Here is what got me. In the same report, that same agent had correctly found the same fault somewhere else in our system. It could name the problem perfectly and still do it.
So now there is a command that checks. Before anybody repeats what an agent says about testing, we count. It takes one second.
Being able to name a fault does not make you immune to it. I have known that about people for thirty years. It turns out to be true about software that talks.
Epics down one side, agents down the other
Picture a grid.
Down one side, the big pieces of work. Down the other, the agents who own a craft.
Every piece of work gets the specialist who has been doing that job all along. Cadence keeps time across the whole grid, making sure nothing goes quiet and everything is still pointed at the goal.
I am not going to sit here and tell you it has doubled our output. It is two days old. Ask me in a month.
What I can tell you is that on Tuesday I could not answer the question “are we closer”, and today I can. That was the actual problem.
The part that is about me
There is a thing I keep noticing and I am going to say it plainly.
The better I get at this, the faster the whole thing goes. Not because I write any of the code. Because a vague instruction produces vague work, and I am the one giving the instructions.
That is a skill. It is not one I had two years ago and it is not one anybody taught me. Knowing what to ask for, knowing when to push back on an answer that sounds right, knowing which decisions are actually mine and which I am just slowing down by holding onto.
We are putting a qualification through awarding body approval that is built around exactly that, and I am taking one myself while we do it. Partly so I can say I have. Mostly because I need it. You cannot ask a company to become fluent in working with these things while being the least fluent person in the building.
The unglamorous truth
Most of this week was not clever.
It was finding out that a button had done nothing for weeks. That a backup had never been tested. That a panel said “no agents running” while six of them had reported that morning, because it was collapsed and I could not see it.
Every single one looked fine.
I have said this before and I will keep saying it. The bugs that cost you a year are not the ones that fall over. They are the ones that work.
One more thing
John sent me a song this week. ‘Til You Can’t, Cody Johnson.
It is a good reminder that however exciting the work gets, and this week it got very exciting, family is always more.