Aeon

Open-source runtime for autonomous agents. Skills you install, schedules you set, a wallet they can spend from.

Comment

Inspiration

Spent a year watching people ship "AI assistants." All of them wait for you to type. Meanwhile I was setting cron jobs by hand every morning to pull the same reports. The question wasn't "can an LLM answer this?" — it was "why is a human still triggering it?"

x402 shipping on Base was the tipping point. If an agent can pay for its own API calls, the human in the loop isn't necessary anymore. That's the whole thesis.

What it does

Aeon is a runtime for scheduled agents. You drop a skill into a folder, add a cron line, point it at a wallet, and walk away. The agent runs its shift, does the work, writes down what it did and what it spent, and shuts down until the next run.

Right now I've got four skills on the clock: a market-data pull, a research summarizer, a social-post drafter, and a wallet-activity monitor. They post to a Telegram bot when they've got something worth showing.

How we built it

Claude Code is the runtime. That call saved months — its skill and subagent primitives already do most of what an agent orchestrator would try to build.

Skills live in git as plain Markdown plus a config file. Cron handles scheduling. No Airflow, no LangGraph, no YAML pipeline. When a job fires, Claude Code loads the skill, runs it, and writes a receipt to disk.

Wallet is a Base agent wallet, funded from my personal one. Paid API calls go through x402 so the agent handles the USDC payment without asking me. Telegram bot is the human-facing surface: status, logs, manual re-triggers.

The whole thing runs on a $12/month DigitalOcean droplet. Same box runs a Paradex trading bot, so I treat it as production and don't push code I haven't tested locally.

Challenges we ran into

The biggest fight was getting the agent to stop wanting to talk. Every LLM tool assumes a human is reading the output. Getting the model to just do the work and file a structured receipt instead of a friendly summary took real prompt work.

Idempotency was next. If the summarizer runs at 9am and again at 10am, it has to remember what it already handled. No chat history to lean on, so every skill keeps its own state on disk.

Then auth. Between xAI, GitHub, X, and the wallet there are five tokens with five expiration policies. One rotated key and a shift silently fails. Lost a weekend of runs to a Grok key I'd rotated and forgotten to redeploy.

Cost attribution was the surprise one. When four agents share a wallet, you can't tell who's burning your USDC unless you tag every request. Ended up writing a small middleware that stamps every call with the skill name.

Accomplishments that we're proud of

The agents run when I'm asleep. That was the whole point, and it works. Woke up last Tuesday to three skills having filed reports overnight, one of them having caught a wallet event I would've missed.

Four skills shipping, not forty half-finished ones. Shipped the boring core before adding a fifth.

The stack fits in my head. Cron, skills, a Telegram bot, a wallet. When something breaks I know exactly where to look.

What we learned

Scheduling is the primitive that turns a chatbot into a worker. Once you see that, half the "agent framework" complexity isn't necessary.

Receipts beat reasoning traces. I don't need to know why the agent did what it did. I need to know what it did, what it cost, and whether it succeeded.

"The agent will figure it out" is a bug. Every time I said it to myself I was avoiding a decision about what the skill should actually do.

Small skills that do one thing beat one big agent that does everything.

What's next for Aeon

A Polymarket monitor and a Dexscreener alerter are the next two skills.

Multi-agent handoff. Right now every skill is independent. Next version: skill A finishes, drops a payload, skill B picks it up on its next run.

A public skill registry. Same idea as npm, but for scheduled agent skills. If someone writes a great X-monitoring skill, everyone else should be able to install it in a line.

Onchain receipts. x402 payment records already form a natural audit trail — I want to make them queryable so anyone can verify what an Aeon agent did and what it spent.

Built With

Share this project:

Updates