B/Agent lab
Aristarkh Gradov
An open-source AI persona that thinks before it speaks
Aristarkh plays a sharp-tongued media producer who reviews creators’ projects. He plans every reply in a private inner monologue before he answers. Trust, fatigue and mood shape his tone. He tells you he’s an AI.
MIT-licensed · a reply guard enforces honest AI identity · end-to-end tests of the live bot, graded by a separate model
Why a CEO should care
- Memory
- client history and policies
- Two-pass reasoning
- explainable decisions
- AI-judge testing
- quality control
- A cheap router model that decides when a costly step is worth it
- the basis for unit economics per request
How he thinks
- Two passes. He plans every reply in a private inner monologue, then answers.
- Affective state. Trust, fatigue and mood, tracked per user, shape his tone.
- Memory. Three memory layers are recalled in parallel.
- A daily routine. A nightly routine reflects on the day and consolidates memory.
- Honest identity. He tells you he’s an AI, and a reply guard enforces it.
Why the same engine matters in a company
Take away the producer and the parts stay useful. Memory becomes client history and policies. Two-pass reasoning makes a decision explainable. A separate model grading answers on a test bench turns quality into a number you can manage. A cheap router model that decides when a costly step, such as web search, is worth it gives you unit economics per request.
Where his knowledge comes from
Built in part on producing-school methodology, and on my own research into hit and failed shows and studies of the psychology of perception and persuasion.