A new chat should not mean a fresh start.
Job applications, documents and project decisions spill across days. I wanted an agent that could remember what happened, retrieve the useful parts and continue from there.
Case study · flagship agent
Henry
The personal work agent I built because useful work should not restart with every chat. Henry remembers the decisions behind a task, finds the right knowledge and runs repeatable workflows from my terminal or phone.
Daily
Job search, documents, standups and scheduled work
900+
Knowledge modules with tested retrieval changes
Local-first
Personal memory and reusable knowledge stay separate
Approval
External actions wait for my explicit decision
why I built Henry.
I was using AI every day, but too much of the work still depended on me rebuilding the context.
An answer is useful. A finished workflow is better.
The repeated work was predictable: find roles, compare them with my resume, prepare documents, collect standups and run reminders. Henry turns those jobs into workflows I can reuse.
The product idea: give one personal agent memory, tools and clear limits. It should carry context forward, do the repeatable work and stop for approval before anything leaves my machine.
how it works.
Every request follows the same path: understand the job, load only the context it needs, choose a focused workflow and return something I can use.
The request path. Terminal and Telegram feed one router. Personal memory and the knowledge library add context before a focused workflow runs.
remember context
Personal history and reusable knowledge stay in separate stores.
route the request
One router selects a capability and the right model tier.
run focused work
Job search, documents, standups and reminders use their own workflows.
ask before acting
External actions stay drafted until I approve the exact item.
how I keep it useful.
A personal agent touches real work, so reliability is part of the product.
- Approval before external actionEmail, applications and other outbound work stay drafted until I approve the exact item.
- Memory with a reasonRecall combines semantic and keyword matches, then shows why a memory was selected.
- Tested knowledge retrievalRanking changes ship only when the evaluation scores improve.
- One phone update loopA single Telegram loop handles incoming messages and prevents duplicate replies.
- Outputs with a job to doResume and cover-letter tools return the finished format instead of a wall of commentary.
Bose turns this engineering pattern into a school product. Henry remains the personal agent I use for my own work.
what's live.
The system runs locally on my 8 GB M1 Air. These are the parts I use as real workflows:
memory
memory that carries work forward
Keeps personal history separate from reusable knowledge, then retrieves only what the current task needs.
knowledge
tested knowledge search
Searches 900+ modules and only ships ranking changes when retrieval evaluations improve.
routing
one request, one clear route
Chooses the right workflow and model tier instead of sending every task through the same prompt.
automation
job scout
Finds roles on Naukri, the open web and X, scores them against my resume and returns a shortlist.
documents
resume and cover tailoring
Tailors resumes and cover letters while keeping the required document format.
telegram
two-way DM bridge
Lets me use Henry from my phone without duplicate replies.
operations
standup and scheduler
Collects standups, sends summaries, runs reminders and shows scheduled work on a dashboard.
publishing
tweet pipeline
Publishes one tech post a day, with a kill switch and a copy sent back to me.