portfolio-press · applied-AI lab
How the work is organised

Agents as a service, not a window.

Read this as a different arrangement of the working day rather than a list of tools: where the work lives, what counts as a unit of work, who writes the code, who reviews it, and what a manager can see without asking anyone. The tools underneath are replaceable. The arrangement is the idea, and it is the part that will look unfamiliar.

The same eight questions, answered twice

On the left, how a development cycle is usually arranged. On the right, how it is arranged here. Neither column is about a product.

the cycle everyone knowsthe cycle on a loop
where the work lives On laptops. It stops when they close. On a server, in named sessions anyone can address.
a unit of work is A ticket, described well enough for a person who will ask questions. A spec, merged as its own pull request, described well enough for something that will not ask.
who writes the code Developers, with an assistant helping them type. The machine. Developers write the specs and approve the result.
who reviews it Colleagues, and increasingly an assistant from the same house as the author. A model from a different vendor, then a person.
where knowledge lives In heads and chat threads. It leaves when people do. In a journal and a written canon the loop reads before it works.
what a manager sees Status reports, assembled by the people being reported on. A queue, and numbers anyone can recount.
what a person is for Typing, and being available to be asked. Deciding. The human is the gate, and only the gate.
coordination Stand-ups, and a team lead holding the state in their memory. One session that writes no code and holds the state on disk.

A day arranged that way

The moments below are taken from real runs and assembled into one day. The day is illustrative; the timings are not invented.

09:10A manager opens three specs. The gate refuses one of them for a figure with no source, and says which figure.
09:12The queue takes the other two. Nobody's laptop is involved; the people they belong to are in meetings.
10:40Three pull requests come back. A reviewer from another vendor blocks one of them; the runner fixes what it named.
11:00An engineer reads the review findings first, approves two and sends one back. That is the whole of their contribution to the code that day, and it is the part that mattered.
16:30The line cuts a release.

and what did not happen

  • No stand-up, because the queue is the status.
  • Nobody asked where a file was, because the canon is written down.
  • Nothing was lost at six in the evening, because nothing lived on a laptop.

The three things this rests on are already published

Rather than restate them here, each links to the page that shows it working.

An agent in an editor window dies with the window

In the editor

The agent's life is the developer's laptop. Close the lid and the context goes with it. Walk into a meeting and the work stops. The second developer cannot see what the first one's agent is doing, and neither can anyone else.

What you get is a productivity feature, one per person, with nothing to measure and nothing to hand over.

On a server

The session runs on a machine that is always on, and the laptop becomes a remote control. Work continues through the night. A colleague can address the same session by name in the morning, and the transcript is on disk either way.

What you get is a shared service with a queue, a review gate and a journal: the thing a number can be put on.

One consequence is worth designing for rather than discovering. A session's transcript exists on disk, and again through the bridge when it is remote-controlled. So a session holding client code never runs on the bridge, and secrets never enter a transcript. That is a rule from the start, not a patch after an incident.

Four questions, not four products

The tempting version of this section is a table scoring the routes on offer. We are not publishing one, for a reason that is the same reason anything here is worth reading: we cannot recount someone else's product, those products change monthly, and being wrong about a competitor reads as spin rather than as a mistake. So here are the four questions instead, answered only where we can prove the answer. Ask them of anything, including this.

Does it run without the laptop?

If the answer is no, everything downstream is per-person: no queue, no handover, nothing to schedule and nothing to measure across a team.

our answer Yes. Sessions live on a server and are addressed by name from anywhere. Fifteen of them were running when this page was written, the youngest twenty-one minutes old and the oldest twenty days. recount: the session list on the host, any day

Who reviews the work, and are they from the same house?

A reviewer built by the same vendor as the author agrees with it. That is not a suspicion; it is measurable, and worth measuring before trusting a review pipeline.

our answer A rival vendor, always. Measured on our own production data: a reviewer from the same model family raised a blocking concern in none of the rounds it saw, a rival vendor in 43% of them. 462 completed reviewer rounds, pulled 2026-08-19

Is there a human on the merge?

Not on the plan, not on the review: on the merge. It is the only place where a person is cheap to insert and expensive to omit.

our answer Every merge, and every deploy. There is no automatic merge anywhere in the loop, by construction rather than by policy. recount: no auto-merge exists to disable

Which of its numbers can you recount yourself?

A dashboard nobody can reproduce is decoration. The test is whether a sceptical person can arrive at your figure without asking you.

our answer All of them. Every figure on this site carries the command or query that produces it, and the strip on the front page is regenerated by a job rather than typed. hover any figure on the solutions page to see its recount

Any substrate can carry this. The loop is the part none of them ship: a spec in, a rival review, a person on the button, and a journal afterwards.

The part I would ask about if I were you

Three failures from these systems, none of them hypothetical.

An updater that failed every two minutes for a month

It rolled back a working tree each time. Half a million log lines, and not one alert. It was found by accident during a cleanup, which is the worst way to find anything.

A pipeline that quietly ran thirty times slower

Hundreds of graphics allocations failed and the work degraded to the processor without an error. A leaking neighbour was holding a fifth of the card.

A session that could not be reached for half a day

It was running the whole time, under a second name nobody could map to the first. Same session, two identities, no way to tell.

Every one of these was caught by a second system with its own memory, never by the one that made the mistake. That is the whole argument for the shape, and the same reason the reviewer comes from another vendor. The weakest block is not the model. It is observability.

What this could mean for a team

Three small steps, each reversible, none of them a platform decision.

step one

Point a second vendor at last month

Read-only, at the agent-assisted pull requests you already merged, and count what comes back. If it is near zero you have learned something worth knowing too.

reversible: nothing is written

step two

Ten small changes as specs

Well described, through the loop, pull requests back. Your engineers keep the merge button. Measure cycle time and review findings against whatever you do today.

reversible: stop after ten

step three

One shared session, named, with a journal

Not an agent per laptop. Learn what a shared agent actually needs from you before anyone has to choose a vendor.

reversible: it is one machine

Everything on this page runs on systems you can look at, with the command to recount each number. If the shape is interesting, the first step is a process review: one of your processes walked end to end, and a written spec you could hand to anyone, including someone other than us.

← prev: the command post you are here · agents as a service next: the trust ladder →