Blog
What we are finding out as we build 1Presence — the experiments that worked, the ones that taught us the most, the decisions behind the product, and what shipped.
There is no fixed test set for a conversation. So we put three models in the room — the agent under test, a simulated user, and a judge — and then spent most of the quarter learning where the judge lies.
We spent a season building the small parts an agent platform is made of, and letting people compose them. Kits are those compositions offered whole — and putting the shelf together told us, twice, what we had actually built.
1 Sep 202624 min
An agent that rewrites its own failing routine also resets what "normal" means for it — so it can never learn whether it helped, and reports the routine healthy either way. The loop built instead measures every repair against how things worked before they broke.
29 Aug 202612 min
Assistants reset; people accumulate. This is the launch-day write-up of 1Presence at full length — the memory economics, the vault, the automation — and the three bets underneath it all: readable memory, a swappable model layer, isolation per person rather than per policy.
27 Aug 202612 min
Between the words you type and the vectors anything searches sits a gap the standard recipe ignores. One question, followed from the moment it is typed to the moment the model answers — and the three places the walk leaves the recipe behind.
26 Aug 202612 min
"Find what I typed", "find what I mean" and "prove it is not there" look like one search problem and are three, mutually incompatible ones. The third arrived by way of an agent asserting a service was in use — from a filename it had never opened.
24 Aug 20268 min
Consent, ambiguity, failure, retry — each quietly assumes someone is present to answer, notice, or click. A schedule removes them all at once, so autonomy here is built backwards from the moment a question has to wait for its human.
10 Aug 20269 min
A dashboard an agent keeps fresh and a person has arranged has two authors, and every refresh must overwrite one of them while preserving the other. We drew that line five times before we saw what it was made of, and it was not between fields at all.
10 Aug 202610 min
Remember everything a person hands over and recall drowns in it; remember only what the agents wrote and their files sit inert while they assume otherwise. Neither default survives — so the decision became the user's, visible, plain-worded, and never made for them silently.
5 Aug 20269 min
Send a message, close the laptop, and the answer should still be waiting on your phone. Getting there meant divorcing the work from the request that started it — then hunting down every place the old weld still showed.
5 Aug 20267 min
The hard files are the ones the assistant did not write: folders arriving in bulk, carrying no signal about what matters inside. What came out is a ladder of four intensities — measured not by how much is stored, but by which machinery a document may touch.
4 Aug 202616 min
A model's window ends; an afternoon's work should not. Sorting out which of two very different overflows actually kills conversations came first — then spilling the blasts, folding what the model sees, and designing for the wall that gets hit at 3am with nobody there.
31 Jul 20268 min
Distance reads as relatedness, size as importance, a line as a relationship — whether or not the layout meant any of it. Making a real memory legible in 3D was mostly the discipline of forbidding accidental marks, on data no demo had prepared.
29 Jul 20267 min
Filed notes, typed facts, documents, summaries, diaries, a conversation index: what looks like one memory is six stores that refuse to merge. The map of the machine — and the evidence that the refusal is the design.
29 Jul 202610 min
Everything an assistant carries into a turn — instructions, identity, two hundred tool schemas — is billed again on every message, for as long as the product lives. So the prompt became a line item: measured with the meter that bills it, laid out for the cache, and gated by a test.
28 Jul 20267 min
A turn's supplier cost is known precisely one moment too late: when it finishes. Between the taxi meter, the truncated answer and the flat-rate subsidy sits this machinery — and the week a credit was made to behave like money.
24 Jul 20268 min
The threat model: your agent reads what strangers wrote. Taking it seriously leads somewhere specific — nothing may depend on a possibly-hijacked model deciding whether it is hijacked, so the guarantees are dumb, deterministic, and parked at the exits.
19 Jul 20268 min
Under-merge and lookups return zero from a store full of the answer; over-merge and two near-namesakes fuse into someone who never existed. Finding the line is a judgment call — so the interesting engineering is everything wrapped around a judgment that will sometimes be wrong.
18 Jul 20269 min
Memory is billed on every message, not once — and the more of it there is, the more there is to bring. Keeping a growing store behind a fixed per-turn spend took layers, caps, and a cache that turned out to be the conversation itself.
14 Jul 202611 min
In a private chat every message is for the agent; in a team channel almost none are. Putting an agent in the room means giving it a teammate's instincts about who is being spoken to: when it has been asked, how long a conversation it was drawn into stays open, and when to stay quiet.
9 Jul 20266 min
A conversational product cannot open a table designer, and prose that stays prose dissolves at the end of every turn. The middle path: recognise the shape as it lands, offer to keep it — and let code, never the model, own the rows from then on.
28 Jun 20267 min
"Stop", "my turn", and "wait — I meant the other one" are three different speech acts, and a greyed-out send button flattens all of them into please-wait. An assistant whose answers stream for minutes cannot take that position.
18 Jun 20267 min
The agent wrote to its memory six times for every time it read from it. A store filed into diligently and consulted almost never is a diary — and no instruction was going to fix that, because the gap was in the architecture.
4 Jun 20268 min
The default architecture for reaching your data is a copy held by someone else. Folder access here was built twice in one evening — and the version that survived reads from your own disk at the moment of the question, and holds nothing in between.
21 May 20268 min
Where does a new specialist come from — a deploy, or a record? The answer decides how fast a platform can grow, and earning it meant splitting the prompt in two: a platform half identical for everyone, and an identity half that is data.
17 May 20267 min
Most software separates customers with a where-clause; an assistant that can act in your name needs a wall its own code cannot cross. One process per person, under its own cloud identity — paid for in cold starts, a provisioning pipeline, and a memory that travels as an archive.
11 May 20268 min
A greeting does not need frontier capability, so we gave the small talk to a cheaper model from another vendor. The experiment lasted an afternoon — and the reasons it failed are worth more than it cost.
2 May 20267 min
What we shipped, what we learned, and the occasional thing that did not work. No drip sequence, no launch countdowns.
Unsubscribe any time. Or take the RSS feed instead.