guard rails
Decides whether an action is safe to take, and stops the dangerous one rather than the alarming sentence.
Judith: personal assistant for group chats.
An assistant for one person is a feature. An assistant in a room full of people is a different problem.
Once an agent is in a conversation with other people in it, ordinary assistant questions become questions about those people. Anyone present can try to steer her. She accumulates things she was told in confidence. And an answer that is perfectly correct can still be the wrong thing to say in front of the wrong person.
Group threads
She replies where more than one person can see her, and anyone in the thread can try to redirect her.
Quietly, in the background
Scheduled check-ins, mail, health signals and a job board that keep running while nobody is asking her anything.
The talk — AI Engineer World's Fair 2026. Watch on YouTube. Starts muted; browsers will not autoplay sound.
Plugins
Judith is not one program. She is a model surrounded by twenty-one small plugins, each responsible for one job and nothing else — one decides whether an action is safe to take, one decides what is worth remembering, one decides whether what she is about to say is fit for a person to see. The ones below carry Sanskrit names, and in each case the name is the job.
smriti
memory
Keeps facts along with where each one came from, and can replace or quarantine any that turn out to be wrong.
sakshi
tracing
Writes down every tool, model and turn, so a failure can be traced to the moment it happened.
avastha
emotional state
Carries how I am doing into what she is about to say.
drishti
observability
Shows every run she made, why any of them failed, and offers to run it again.
kavacha
armour
Maps where I am to how careful she should be.
collection
Brings in news, filings and market data, and only ever reads them.
upaya
remedy
Matches a repeating error to the fix that worked the last time it appeared.
sarathi
model router
Sorts a request into a lane and hands it to a separate session to answer.
switched off
21plugins around the model, each with a single job
77,919steps recorded across 57 days of continuous operation
65times the last check caught something before it reached a person