On 15 September 2026 the operator of this estate put a name to the way they work with the agent: “You might call it: just-in-time information, with parked placeholders.”
Each half came with its own sentence. The first: “There is no need for me to spill my guts about something until we are ready to work on it.” The second: “But I do need you to capture things so I remember to come back to them later.”
Hold back what the task doesn’t need
The operator knows far more about this estate than the agent is told at any moment: why a decision went the way it did, what failed somewhere else years ago, which parked item is about to matter. That is deliberate, and the reason is not secrecy.
Holding back is more about keeping the context relevant to the task at hand.
An agent can only usefully consider so much at once. The operator, in their words, retains “much more context than what is possible for the agent to consider, without trying to boil the ocean.” So each session gets what its task needs, and nothing it would have to wade through.
We had this wrong at first. The agent’s account said the point was to keep its own recommendations uninfluenced, so that the operator’s agreement would confirm something. That does happen, because an agent that hasn’t been told the answer has to reach its own. But it is a side effect, and the operator corrected the account.
The documents work the same way. The file every session reads first points to where a detail lives instead of repeating it. Its index of known traps is one line each, so an agent knows a trap exists, and the full entry gets read when someone touches that part of the system. Orientation is paid for at the start of every session, so it carries only what every session needs.
Keep a place for everything else
Holding back is only safe if what’s held back isn’t lost. That is the second half of the name, and most of it is the agent’s job.
Each side covers the other’s forgetting. The agent forgets everything between sessions, so the documents hold what was agreed. The operator can’t hold every open thread, so the agent keeps the placeholders: a parked item that names what brings it back, a doubt that says what would settle it, a one-line index entry that points at the full story.
Where each placeholder came from
None of these were invented for the agent. The operator brought them from elsewhere and gave them names an agent could act on.
- The parking lot comes from interviewing users for business requirements: “Most users will tell you more about their pet peeve than they will about core business functionality, so you have to capture their concerns, and have a place to put them where they don’t feel ignored. Same is true for an agent.”
- Doubts is a word from Brazilian developers: “That was the word they used to express their questions. I thought it was a perfect way to frame it, and have been using that word ever since!”
- The settled list, decisions marked closed, came from working with agents, because “new sessions would keep revisiting issues that needed to be deferred.” It has a gentler twin, deferred, not dropped: “a kinder, gentler way of saying the same thing, because ‘DECIDED — do not reopen’ was sometimes a way to get the agent to let it go for the session.”
- Gates, conditions that must hold before work starts, come from everyday life: “Gates are everywhere in life, but we usually don’t label them as such. This was a way to communicate a logic condition to the agent that was easily understood by both of us.”
- One push-back, then drop it, where the agent states an objection once and then lets the decision stand, the operator can’t place yet: “Not sure, we can come back to this with more context.” So it has a placeholder too.
What we measured, and what we couldn’t
A method built on placeholders can fail quietly. Things get parked and never return, and the parking lot becomes an archive. We wrote that down as a doubt about our own method: perhaps deferred, not dropped is just a well-kept backlog.
It was settled on 15 September, by testimony rather than a count. Asked directly, the operator said: “Yes, parked items are constantly coming back. A parked item usually does carry weight.” Nobody has tallied how many returned because their condition fired.
The same day, the agent raised a sharper doubt: that no doubt in the register had ever been closed. It measured that from the file’s history. There were ten doubts, and none had ever been deleted, which is the rule, because a doubt is closed by writing down what settled it. But the measurement was of the wrong store. Most doubts are raised in a session and settled there, and never get written down as doubts at all. The agent withdrew its doubt.
Most doubts never reach the register. They are raised in a session and settled there.
The operator was frank about both halves of that. Asking for the agent’s doubts “keeps you from nagging me”, and most of them “usually don’t really hold any weight.” But also: “Some of your doubts turn into real work — you are a great help in thinking of things that I haven’t thought of.” A parked item usually carries weight, and a doubt raised in a session usually doesn’t. That is why the first gets a written condition and the second gets heard, and most often dismissed.
Where it fails
This method doesn’t fail when information is missing. It fails when a placeholder is.
We retired a key and wrote down that every machine refused it. Nothing pointed at the one place nobody had checked, so no session went looking, and on 4 September we found it was still the Git server’s administrator key, two weeks after we had declared it gone. That story is The key that was retired everywhere.
What it cost
- Two weeks of a live administrator key, because no placeholder pointed at where it was registered.
- Two wrong accounts of this method, both the agent’s. It described the practices as having grown out of its corrections, when they are the operator’s own. And it described holding back as a way to keep its answers uninfluenced, when the point is relevance.
- A doubt measured in the wrong place, raised and withdrawn on the same day.
The rule it bought
Give the agent what the task needs, and give everything else a placeholder. Context that isn’t relevant yet doesn’t help the agent. It just has to wade through it. But every thread you hold back needs a slot that says what brings it back, kept by whichever side won’t forget it. With an agent, that side is the documents.
Where we might be wrong
That parked items come back is the operator’s observation, not a measurement. Nobody has counted how many returned because their condition fired, how many were found by accident, and how many never returned. If most come back by accident, the named condition is decoration, and this is a well-kept backlog after all.
And all of it comes from one estate, with one person and one model. Holding back works because that one person knows what they are holding. On a team the context isn’t all in one head, and deciding what each session needs becomes a coordination problem this estate has never had.
None of this is new
The operator’s own long view comes from two books they have read, The Goal and The Phoenix Project, and the lesson they took from both: “We find the bottlenecks and fix those first.”
In The Phoenix Project the bottleneck is a person. Brent is the engineer every problem waits for. In the publisher’s summary: “Brent’s expertise becomes a bottleneck, as he is constantly pulled into firefighting mode, leaving little time for proactive work on initiatives like the Phoenix Project.” What he knows is in his head, so the work queues behind him.
Eliyahu Goldratt wrote The Goal, and The Phoenix Project is built on his work. Writing about Henry Ford’s flow lines, he named the counter-intuitive part: a factory needs “a practical mechanism that guides the operation when not to produce (prevents overproduction)”. Toyota took it further and changed how industry sees stock, “from an asset to a liability”. In Goldratt’s own scheduling, the bottleneck sets the pace and work is held back before it starts, “the action of choking the release”.
With an agent, the bottleneck is its context. It has room for only so much, and anything loaded before the work needs it is stock: it takes up the room and hides what matters. Holding back context is choking the release. The placeholder is what Brent never had: what isn’t released yet is written down where the next session will find it, not kept in anyone’s head.