An agent workspace
The jobs your team does by hand every week could be running on their own.
It is a thing that happens: a summary already written when you sit down in the
morning, an invoice read and filed the moment it arrives, a question answered out
of documents nobody wants to open again, a job that runs overnight and only wakes
somebody if something looks wrong.
pokrok is where those live. You get an instance of your own, on your own address,
with the model connection already included: no provider account to open, no keys
to paste, no second bill to work out. Nothing to install and nothing to keep
running.
Free while this is in preview. After that, one price for the whole instance with the model connection included, and no seats to count.
Recorded in a running instance.
Monday morning
The report somebody
rebuilds by hand is finished before anyone is in.
The inbox
Read, labelled and filed
before you open it, with only the ones that need a person left for you.
Three weeks later
Ask what you
decided in March and it knows, because it was in the conversation.
What this takes off your week
Everybody has three or four of these: the report somebody rebuilds every Monday,
the inbox somebody sorts by hand, the question somebody answers for the fifth
time. They are the things that keep eating a Thursday, and they are the things
that stop.
The same job, every week, forever
You already know the steps. Doing them is the boring part
Say it the way you would explain it to a new starter: check these pages every
hour, and tell me only if something is not normal. An agent builds that as a
workflow and shows you the steps it built. From then on it goes on the clock,
or the moment the work turns up: an email, a message in Telegram, Slack or
Discord, a push to a repository, a new item in a feed, a file dropped in a
watched folder, or a call from another system.
Shown: a workflow running, a step at a time.
Answering the same question again
Everything your team knows is in somebody's head or somebody's inbox
It was in the room when you decided. Why you changed supplier in March, the customer who always pays late, what was settled in a chat in January: it picks those up as they happen and still knows in September, whether somebody asks in a chat, in a Telegram bot, or inside a job running at three in the morning. Every answer carries the passage it came from, so anybody can check it.
Shown: a question answered from one team's own notes, with the line it came from.
The assistant you cannot quite trust with anything
Give each job its own agent, with only the access that job needs
The one that answers customers can read the handbook and the support inbox and nothing else. The one that reviews code can reach the repository and not the inbox. The one that handles invoices can write to the finance folder and cannot send mail in your name. You decide that for each of them, and you can see on one screen exactly what each one is allowed to touch. One can also hand a job to another and wait for the answer.
Shown: an agent, and exactly what it is allowed to touch.
The task nobody automates because it needs a machine
Some jobs need a real machine, not an answer
Converting a folder of files, rebuilding a report, running the tests on a branch before anybody reads it: these are the ones that never get automated. Each instance gets containers of its own, so an agent installs what it needs, runs the job, keeps what it installed for next time and hands back the result, or a web address if what it built should be looked at. It can reach the internet when a job needs to, and you can switch that off for one that should not.
Shown: an agent installing what it needs, running the job, and handing back the result.
The cookbook has the exact words to ask for each one, so
somebody on your team can copy one and have it running this afternoon.
What makes it different
Memory
It remembers what happened, not just what you filed
Ask in three weeks what you decided and it knows, because it was there. The
conversations it takes part in are kept, and the parts worth keeping become
things it uses later, so nobody explains the business to it twice. Searching
a document you handed it is the least of what it does.
Automation
Say what you want to happen. It writes the workflow and shows you the steps
Fetch a page, run a command, ask an agent, branch on the answer, post the
result, wired together into something that runs without you. You do not
assemble it: you describe it, an agent builds it, and it tells you what is
still missing. If a step fails once it is running, the run stops there and
keeps the error, so you can see which step and why.
Agents
Hand an agent as much or as little as you like
Give an agent the run of your workflows and it will build them for you.
Give it a machine and it will set one up. Describe what you want in a sentence and
it makes it, then tells you what is still unconnected.
Its own machine
Somewhere to actually do the work
An agent can install something, run a build, break it and try again -
inside a container that belongs to you. Whatever happens in there stays
in there, and a bad afternoon costs you a container rather than a
machine.
Seeing what it did
When an answer looks wrong, you can see why, and what each job cost
Every request and reply is kept exactly as it went: what was asked, what came
back, how long it took and what it cost, job by job. No summary standing
between you and what actually happened, and no surprise at the end of the
month.
Yours
Nobody else is in your workspace
Your instance is yours, on your own address. Your data is not pooled with
anybody else's. Enterprise customers who need it on their own infrastructure,
because of a policy or a jurisdiction, can have it there instead.
Everything it does →
Bring your own machine
An agent can work in a container here, or on a machine you already own: a build
server, a box with a GPU, your own laptop. Paste one line onto it and it joins
your workspace like any other machine an agent can use. Nothing on it is opened
to the internet, because the machine dials out to your instance rather than
waiting to be called.
How that works
Start with one routine
Pick the job somebody on your team does every week without thinking about it.
Describe it in a sentence, watch it get built, and see whether next Monday is
different. The cookbook has the words for the most
common ones.
Free while this is in preview. After that, one price for the whole instance with the model connection included, and no seats to count.