AgentDesk
PlannedAn assistant that proposes an action, but always waits for a human to approve it.
An assistant that diagnoses a customer problem, proposes an action (a refund, a handover to a person) and waits for someone to agree before doing anything.
The problem
An autonomous assistant with money on the line is an incident waiting its turn. The question is not "can it do the job", it is "what stops it from doing anything at all": a budget, a scope of actions, a human stopping point, and a record of what was decided.
How it is built
- 01
A reasoning loop written by hand, then rebuilt with a dedicated tool that saves its state.
- 02
Triagely and DocPilot made available as tools: the three projects talk to each other.
- 03
Strict spending and duration caps, with a clean stop when either is exceeded.
- 04
No consequential action without one person explicitly approving it.
- 05
An interrupted conversation resumes exactly where it left off.
- 06
The assistant is also usable from an external chat tool, with authentication.
- 07
Answers shown as they are written, with no needless waiting.
What it proves
A bounded assistant: capped budget, human approval, resumption after interruption.
Build log
The full write-up will land here: what was tried, what failed, the numbers obtained, and what I would do differently. It goes up once the measurements exist, not before.
Triagely