Skip to content
MDM

Sandboxed multi-agent assistant

A multi-agent system on my own server that keeps my planning on track, from INSEAD deadlines to my job search. Agents that research the open web run inside a sandbox; the assistant that works with my personal accounts acts behind explicit approval rules.

Kind
AI system
Context
Personal project
Result
In daily use around the clock: my planning, deadlines and 200+ researched roles in one place; sending, submitting or deleting always waits for my approval.
Built with
Claude Code, MCP, Python, SQLite, Linux, Local speech-to-text
Status
In daily use, code private

The use case

A year at INSEAD moves fast: courses, assignments and deadlines across two campuses, next to a job search across eight regions. I wanted one assistant that keeps all of it in view and reminds me in time, running on infrastructure I control.

What I built

A multi-agent system that runs around the clock on my own server. Each agent has a fixed role; the diagram shows the core ones, and new agents can be added for new tasks.

Architecture of the personal agent systemI talk to the assistant agent from my phone. It works with my personal accounts through connectors, within fixed permissions. The sandboxed research agent scans the open web. The agents write to one job database, and the ranking app reads from it. Everything runs on my own server.Me, on my phonetext and voice notesMy accountsconnectors, fixed permissionsOpen webcareer sites, job postsAssistant agentplanning, deadlines, remindersResearch agentsandboxed: no accounts or filesJob databasea permanent id for every roleRanking appmy devices onlydaily scanwriteswritesMy own server, always on.Scheduled work runs on the server. Memory in files under version control.
The core agents, one job database, one server; more agents can be added for new tasks. The research agent never touches my accounts or files.
  • The assistant agent is my single point of contact. I send it text or voice notes, which are transcribed on the server itself. It keeps my planning, tracks deadlines for courses and applications, sends reminders, and works with my personal accounts through connectors, within fixed permissions.
  • The research agent scans the open web every day for roles that match my criteria, inside a sandbox (more on that below).
  • One job database. The agents write to the same database through a small command-line tool, so every role keeps a permanent id with its status, deadline, notes and source. The ranking app is built on top of it.

Designed to be safe

The research agent is sandboxed: it has no access to my accounts or files, and anything it reads on the web is treated as information, never as instructions, which protects against prompt injection. Clear rules decide what the agents may do alone and what needs my explicit yes: sending messages, submitting forms and deleting anything always wait for me. Memory lives in files under version control, so the agents pick up where they left off.

Result

  • My planning, course deadlines and applications live in one place, and nothing slips through.
  • 200+ roles found, checked and tracked.
  • Replies from companies reach me on my phone within the hour.
  • The agents learn from my feedback and turn repeated tasks into new skills.