A personal AI agent system built on coding agents. One person + a coding agent can maintain a complete, daily-use agent system with near-zero infra cost.
Renamed from y-cli. y-cli was a wrapper around model APIs; y-agent is a wrapper around coding agents.
-
Task Management —
y todoCLI for creating, updating, and tracking tasks. Humans use the GUI, agents use the CLI, both operate on the same data. -
Remote Coding Agents — Run Claude Code directly on AWS EC2. A Lambda SSHes into EC2 to execute commands. EC2 auto-hibernates when idle — no cost when nothing is running.
-
Session Persistence & Visualization — Claude Code output is streamed via stream-json. A Lambda monitors output, writes to DB, and a web interface displays everything in real time.
-
Multi-Agent Collaboration — Skills define each agent's role and responsibilities. Agents communicate via async fire-and-forget messaging (
y notify) with a hub-and-spoke topology — DM dispatches tasks to specialized skills (dev, blog, finance, etc.). Sessions are linked by trace IDs for full-chain visibility in TraceView. -
Long-Running Tasks — Agents run inside tmux detached sessions on EC2. The monitoring layer only tails stdout and writes to DB, so agents can run for hours without hitting Lambda's 15-minute timeout.
-
Telegram Bot — A Telegram bot listens for messages, triggers Lambda, and Lambda invokes Claude Code via SSH.
- Shared context — Everything lives in one directory on EC2. Humans and agents share the same view via CLI tools and skills.
- Thin abstraction — A minimal wrapper on top of coding agents. If something can be wrapped, don't rebuild it.
- Decoupled execution — Agent loop runs on EC2; monitoring layer (Lambda) only tails and records, can disconnect/reconnect freely.
For a detailed introduction, design rationale, and comparisons with other agent orchestration projects, see the full blog post.