← What I've built

Agent fleet

A team of AI agents I manage the way you would manage people: clear lanes, the right seniority for each job, and a rule that every agent reports back.

Why

One person, many projects

I run a job, a trading system, a life-tracking app, research for my dad and a home server. No single conversation with an AI can hold all of that, and a single long session gets slow and expensive. So I split the work into lanes. Each lane is a separate agent session with its own folder, its own memory and one job.

How it works

Lanes, tiers and a chain of command

  • Me
  • Orchestrator
  • Project architect
  • Builder lane
  • Reviewer lane

An orchestrator agent knows every project and spawns lanes when work needs a dedicated mind. Each lane is a terminal session I can attach to from my phone. Lanes are named after what they do, so the fleet reads like an org chart.

The most capable model acts as the manager of a project: it keeps the architecture and judgement work and hands everything routine to cheaper models. A monitoring lane runs on a mid-tier model. Deep research gets the strongest model. Every spawned lane is given a tier, with one line explaining why.

Model tiersSmall, mid, large and research-grade models, each priced and assigned on purpose.
ReportingA lane that finishes or gets blocked messages its parent immediately. A timer is only the fallback.
Usage budgetEvery agent sees the shared plan usage, estimates a task before starting, and measures the real cost after.
Hard linesAnything touching real money is changed by the senior lanes only. Monitoring lanes watch and escalate.
What is interesting

The problems are management problems

  • Cost is a design constraint. A cheap call that needs three retries is not cheap. Lanes escalate early instead of struggling.
  • Context is the scarce resource. Handoff documents let a lane clear its context and resume cleanly instead of dragging a huge history along.
  • Trust, then verify. Work is not done until it is merged, reported and checked by something that is not the agent that wrote it.
  • The box is disposable. Everything durable lives in GitHub plus an encrypted secrets backup, so a dead machine can be rebuilt from scratch. A push gate scans for secrets before anything leaves.
  • Human attention is rationed. When an agent is blocked on me it sends a phone notification with a one-tap link. Otherwise it keeps going.
Stack

What it runs on

Claude Code, tmux, git, shell scripts, systemd timers and a small set of hooks and skills I wrote so agents can wake each other, hand off work and publish pages. A shared gallery collects every report the agents produce.