Felix
Felix is the layer between what you said and what gets done.
An ultralight voice-to-outcome agent harness. It carries a spoken request all the way to the action without rewriting it on the way.
In development · not released
One request, all the way through
You said
“Push tomorrow’s onsite to next week and tell whoever’s flying in — not the ones dialling in.”
Felix pinned down
- tomorrow’s onsiteThe Acme onsite, Thursday 27 AugustResolved against the day it was said. The same four words meant a different meeting yesterday.
- next weekThursday 3 September, same slotSame weekday, same hour. That is what you meant, and it is the part nobody bothers to say.
- whoever’s flying inThree of the seven invitedThe three with travel booked against that date — not a guess about who usually travels.
- not the ones dialling inFour people, deliberately not toldA negation. It is the first span a fluent rewrite drops, and the only one here you would notice at the meeting.
Then showed you this, and waited
- Move the invite to 3 September
- Flag 3 flights that no longer fit
- Message those 3, not the other 4
Every phrase in the left column is in the audio, unedited. Felix resolved what those phrases pointed at. It did not decide what you meant.
The name, taken apart
Voice-to-outcome
The unit is the outcome, not the transcript. A flawless transcription of a request that then runs wrong is a failure — and it is the failure most of the stack is currently optimised for.
Ultralight
Not a platform, not a place to work, nothing to migrate into. It sits between the microphone and the tools the work already lives in, and it is meant to be forgettable.
Agent harness
A harness is what holds a working animal to the load. Felix does not replace the agent. It holds one to what you actually said.
What it holds on to
Four things in a spoken request decide what actually happens: who and what you pointed at, the numbers and times, the conditions, and the order they run in. Felix marks those and carries them through unchanged. Nothing in the pipeline is allowed to paraphrase a request, because the rewrite that reads best is the one that costs most — we wrote up the measurement rather than assert it.
What it will not do
It will not tidy your speech. Cleaning a spoken request into careful written English was the most damaging operation in the study we build on: twenty-four points of downstream accuracy, gone. Deleting the ums recovers nothing at all. The disfluency is not the problem; the tidying is.
It will not guess when it is unsure. An agent that asks finished 69% of underspecified requests, against 71% for one handed the full specification. Asking costs almost nothing. Guessing costs the outcome, and you find out later.
It will not score itself on the transcript. The numbers we hold ourselves to are how much of what we produce you have to correct by hand, and whether the thing that got done was the thing you meant. Word error rate moves without either of those moving, which is why we stopped using it.
Where it is
Felix is not finished, and we would rather say so than stage a demo of it. It is being built with a small number of teams whose work is already spoken out loud — where there is no free hand for a keyboard, or where the conversation is the work. There is no date, and we are not going to invent one. The reasoning behind all of it is on the thesis, with its sources.
Early access
Tell us what you would point it at.
We are adding teams one at a time, because each one changes what Felix has to hold on to. It is a conversation before it is an account — a person reads what you send, and a person answers it.
Nothing to install yet. You are telling us what to build, not signing up for something.