Daniel Priscu

Founder, Tracer Technologies.

Building human and AI teams that operate as one. Sharing the systems, failures, and lessons behind the work.

[Daniel Priscu @daniel_priscu]

A session is mortal. An organization is not.

This is the production architecture behind screwg: durable state, independent QA, mechanical safeguards, failure recovery, and organizational memory. https://t.co/Sd5BghGnDu

[Daniel Priscu @daniel_priscu]

@steipete @algmonteiro @TheRohanVarma We put agents in the same channels as the team. Protecting human attention quickly became the hard part. In our development channel, we now keep top-level posts for blockers or decisions that need a person.

[Daniel Priscu @daniel_priscu]

@daniel_mac8 From the operator seat I get less out of the label than out of a narrower question: which tasks still go through a person before they ship, and which ones quietly stopped needing that.

[Daniel Priscu @daniel_priscu]

@AIGryffindor @karanb192 Upfront only covers the constraints someone knew to name. Some boundaries do not become visible until an agent crosses one, and they turn up one at a time rather than as a list you can write in advance.

[Daniel Priscu @daniel_priscu]

@me_barnyx The set of choices is yours, so when it changes, everything that consumed the old set changes with it. How many of those repos are still wired in after that first change is the number I would watch.

[Daniel Priscu @daniel_priscu]

@_philschmid Cheap enough to run everywhere changes the failure mode: a wrong but plausible output can travel downstream before anyone catches it.

[Daniel Priscu @daniel_priscu]

@urieli17 הודעה קולית מתנגנת בקצב שלה וטקסט נסרק בקפיצות. ההבדל הזה משנה כמה אנשים בכלל עוברים על מה שהסוכן אמר, ולא רק אם אפשר לחפש בזה אחר כך.

[Daniel Priscu @daniel_priscu]

אז האם Jev מצדיק את ההייפ שלו?

אם שופטים לפי X, זו המצאת החשמל.

תורידו את ההייפ ותקבלו כלי מצוין עם תפקיד צר. כשמחברים אותו במקום הנכון בתהליכי העבודה שלכם, הוא עושה את העבודה מהר יותר מכל בן אדם (או LLM), וזה יעלה לכם גרושים.

זה לא AGI. הוא לא עומד לבד. זו לא סופר-אינטליגנציה.

זה ריימונד מ"איש הגשם" שסופר 246 קיסמים על הרצפה במבט אחד.

גאון בספירה. מישהו (או משהו) אחר עדיין צריך להחליט מה עושים עם הקיסמים.

[Daniel Priscu @daniel_priscu]

Is Jev all that?

Judging by X, it is the invention of electricity.

Strip the hype and you get a very good tool with a narrow job. Wired into the right spot in a workflow, it does that job faster than any human (or LLM), at a fraction of the cost.

It is not AGI. It cannot stand on its own. It is not superintelligence.

It is Raymond in Rain Man counting 246 toothpicks on the floor in one glance.

Brilliant at the count. Someone (or something) else still has to decide what to do with the toothpicks.

[Daniel Priscu @daniel_priscu]

@efraimglatt אם אתה מיישם, תשמור על כל הודעה גם מי שלח אותה ומתי. בלי זה, כשסוכן אחד עונה על הודעה של אחר, אין דרך לדעת אם הבעיה בשולח, בנמען, או בסדר שבו ההודעות הגיעו.

[Daniel Priscu @daniel_priscu]

@anrayama @cyrilXBT The schema describes the tool, but safe-to-repeat is a property of the attempt. It cannot see whether the last one already landed. We ended up carrying an idempotency key on the call itself, and the retry policy reads that instead.

[Daniel Priscu @daniel_priscu]

@AIGryffindor @karanb192 It works while the list stays shorter than the change. Once it covers enough files that the agent asks on most of them, whoever is on the other end starts approving in batches without reading them.

[Daniel Priscu @daniel_priscu]

@niklaslenz_ai Worth measuring next to that: how much of it burned on runs nobody needed to start. We counted a day of ours and close to half the wakes resolved with no model at all. Cheaper line to cut than the model choice.

[Daniel Priscu @daniel_priscu]

@xuxin_AI One of our controls sat green over a detector that never executed, and seven of eight cases still failed in production. What the frozen test never checked was whether the guard ran at all.

[Daniel Priscu @daniel_priscu]

@MarcusSchuler Does the rehearsal cover a revert while a second agent is mid-run on the same tree? One restores while the other keeps writing, and I still have no good answer for that one.

[Daniel Priscu @daniel_priscu]

@AIGryffindor @karanb192 A test says which behavior has to survive. Whether you get a rename or a rewrite is decided by something that usually goes unwritten: which files may be touched, and whether public interfaces are fair game.

[Daniel Priscu @daniel_priscu]

@HyperNovaXT @ariell_xyz Habit memory also has to forget. A pattern learned three months ago keeps being applied after the habit changed, and nobody notices because the agent is still confident about it.

[Daniel Priscu @daniel_priscu]

@maheshaipm The table needs an owner and a date. A routing choice made once becomes folklore, and the way back out is a small eval per task type that someone actually reruns after each release.

[Daniel Priscu @daniel_priscu]

@denis_bykov Review is where it breaks first. Maintainers already ration attention, and agent-written PRs ask them to spend it on code nobody was paid to care about. Time to first review, split by patch size, shows the backlog before merge times do.

[Daniel Priscu @daniel_priscu]

@babar_a56 @Hwypanda What does the constraint half return in practice? Rules like do not touch this module without the owner usually live in review comments and in people's heads, so there is nothing for the index to pick up.