AI Systems
TARS: an Agentic AI Operations System
A personal AI chief of staff with a 1,900-node knowledge graph, live control of my infrastructure, and standing multi-agent workflows.
- Knowledge graph
- 1,900+ nodes
- Graph edges
- 4,700+
- Automations
- Daily
Problem
AI assistants forget everything, control nothing, and wait to be asked. I wanted the actual TARS: an AI system with durable memory of my life and systems, real operational control of my infrastructure, and the ability to run work on its own schedule, not just answer questions in a chat window.
Approach
TARS is built on three layers. The first is memory: a version-controlled knowledge vault that functions as a second brain, compiled into a knowledge graph of more than 1,900 nodes and 4,700 edges covering my systems, projects, decisions, and history, with semantic retrieval on top so the AI recalls precisely instead of generically. Every non-obvious call gets written down as a decision record, so the system can tell me not just what it did but why it was done that way.
The second is control. Through a set of connector integrations, TARS operates the infrastructure I run: the hypervisor, the network edge, DNS, personal email, home automation, workflow engines, and the machines themselves over SSH. Asking it to do something means it does the thing, not that it tells me how. Employer systems are deliberately outside that boundary, and keeping the line clean is part of the design rather than an afterthought.
The third is autonomy. TARS runs on a single always-on head machine that holds the live state, with every other machine in the lab attaching to it as a client and pull-only mirrors standing by, so the whole system can move to different hardware in under an hour. Scheduled automations mine my email into the knowledge base, brief me every morning, and sync my task list. A standing multi-agent workflow puts multiple AI models from different vendors in defined seats with defined roles, reviewing and verifying each other’s work.
Result
A system that compounds: every day it runs, it knows more, and everything I automate stays automated. It is the most complete expression of how I think about engineering, where memory, control, and autonomy turn a chat model into an operations partner. It is also, frankly, the project I would show anyone who wants to know what I can build when nobody hands me a spec.