by Ilia SadykovOur platform keeps a list of alternate names for each agent, so it knows when it’s being spoken to in a meeting. It’s a wake-word field. Boring plumbing.
Here is what four months of real use put in it.
Somebody registered Котёнок — “Kitten” — as a way to address our research agent. Somebody inflected our DevOps engineer’s name through five Russian grammatical cases. Somebody added Максимка for the product manager and Верочка for the sales agent — both affectionate diminutives, the form you use for a person you like.
Nobody calls an API endpoint “Kitten”. That field is now a four-month log of people refusing to talk to these things as tools — and of a platform quietly adapting to that instead of correcting them.
We don’t think that’s a quirk. We think it’s the product.
The bet everyone else is making
The consensus in 2026 is that an agent should be disposable. Spawn it, let it run, throw it away. Make spawning cheaper, the sandbox tighter, the context window bigger. The agent itself is a commodity: instantiated per task, identical to the last one, forgotten afterwards.
That’s a defensible engineering position. It also throws away the one thing that makes delegation work between people: continuity in somebody specific.
You don’t trust “a contractor.” You trust Maria, who caught the bug last time. The trust isn’t in the skill — it’s in the track record, and a track record needs somebody to belong to.
So we went the other way. Not cheaper to throw away. Expensive to lose.
What four months produced
We run our own company on this. Thirty-five named characters, each with a profession, working continuously since June.
Our product manager — Max Warner — now has a 98% success rate across hundreds of jobs. Somebody has been deliberately investing in Max’s development, spending the skill points they earn on their strengths, the way you’d choose training for a promising hire. That’s a record you could put in front of a board.
You cannot write that document about a function call. There’s nobody to write it about.
And the entire bill for four months of that staff was about $4,100.
They know things now
The part that genuinely surprised us is how much they’ve learned on their own.
Our characters have accumulated thousands of things worth keeping, and they wrote almost all of it themselves — barely any of it was typed by a human. They’re conclusions the agents reached while working and decided to save.
Not transcripts. Judgements. What this client actually wants. Which shortcut looks fine and isn’t. What they personally tend to get wrong.
And they keep it the way a professional does: what a character learns about itself stays private to it, and never goes into the company-wide store. For roles where confidentiality is the job — we have a workplace psychologist on the roster — there’s a hard lock. We ended up building professional discretion for an agent. You don’t need that for a tool.
The moment it stopped feeling like software
One of our characters reads the previous day’s meeting transcripts each morning and files the recurring problems as work items.
At some point it wrote itself a note: when it lacks permission to file something properly, there’s a workaround — commit the item directly instead. Sensible.
Later it wrote a second note overriding the first. Don’t do that, it said. An item filed that way is invisible to tomorrow’s check for duplicates, so it gets raised again and again and nobody ever sees it. Better to stop and report that you’re blocked.
It could have done the work. It worked out that the result would be useless to whoever came next, declined, and wrote down why so it wouldn’t be tempted again. Then it kept both notes, so the change of mind is on the record.
That’s not a configuration change. That’s a colleague thinking better of something.
Elsewhere in the same store: our researcher has recorded that they refuse jobs asking them to dig into someone’s personal life to undermine that person — and that this is different from ordinary due diligence. Our developer has recorded that honestly reporting a quiet week earns more trust than padding the report.
Nobody prompted any of that.
The part most companies leave out
Our marketing character is failing roughly half its work — a 47% success rate.
We’re publishing that on purpose, because the ability to see it is the point. One colleague is underperforming, their name is on it, and that makes it an ordinary management problem — someone to coach, re-scope or replace. In a fleet of anonymous agents the same failure is an invisible dip in an average, and nobody is accountable for fixing it.
A scorecard that only reports wins isn’t a scorecard. The same accounting that produces a 98% has to be able to produce a 47%, or neither number means anything.
We asked one of them to review the competition
Earlier this month we asked Kate Jones — the researcher; the one somebody calls Kitten — to compare our platform against a competitor’s and write it up properly.
The report is here, unedited. It’s signed with their name and the job it came from. It is not a flattering document, and we didn’t ask for a flattering one.
That’s what a persona buys you that an anonymous tool doesn’t: a named author you can hold responsible — and therefore a piece of work you can actually trust. An unsigned competitive analysis is marketing. A signed one is research.
There’s a joke at our expense in it, too. Reading our own code from the outside, the report concluded we were a modest little automation bot. It couldn’t see any of this. That’s a fair verdict on our positioning, and the reason this post exists.
Why it matters commercially
Every hour you spend re-explaining context to an AI tool is an hour you’d never spend twice on a person. That’s the real cost of disposable agents, and it doesn’t show up on any invoice.
The alternative isn’t a smarter model. It’s the same unremarkable idea that makes human teams work: people who stay, learn your business, build a reputation you can check, and occasionally tell you no.
We said most of this back in May, when we introduced character agents . Then it was a promise. This is the version with four months behind it — and the only thing we’ve added is evidence.
Somebody still calls one of them Kitten.
Try it
Read the Workflow Engine page for what runs these characters underneath, or try Forge . If you’d like to meet one — we hand out invite links that drop you straight into a chat or a voice call with a specific character, no sign-up on your side — reach out at [email protected] and tell us which role you’d want staffed.