How Penchant was conceived
Penchant did not start as a product. It started as a workload.
The person building it runs a small company and, like a lot of people, hit the point where the admin around the work had quietly become most of the work. The obvious move in 2026 is to throw agents at it. The problem is that the moment you let software act on your behalf — send the email, move the money, change the record — you have handed it two things at once: the tedium you wanted gone, and the decisions you did not.
Most tools blur those together. An agent does forty things and tells you about them in a chat log you skim and trust. That is fine right up until the one action you would never have approved is buried on line thirty-one, phrased exactly like the thirty you would have.
Penchant is a bet that those two things should come apart. Agents should carry the tedium completely and carry the consequential decisions not at all — those should arrive back at a human, prepared, with the evidence attached, and wait.
That splits the whole system along one line, and everything else follows from it:
- The work the agents do is recorded in a ledger you can actually read, not narrated in a chat you have to police.
- Anything consequential stops and asks a person, and it cannot talk itself past that stop.
- The memory the agents reason over cites where each thing came from, and it cannot quietly rewrite its own history to make a past action look justified.
What it sets out to achieve is small to say and hard to build: a day of work arrives done, with only the decisions that genuinely needed you routed back to you.
None of that is for sale yet, and this log exists partly to keep us honest about the distance between that sentence and what actually runs today. The next entries are the specific decisions behind each of those three commitments — starting with the front door.