Workshop notes
Field notes on running autonomous agents
I build AI agents that trade, sign transactions and maintain websites without asking me first. Here I write down how they are built, what holds up in daily operation and what breaks.
Chainwhizz is my workshop. Everything described here runs on my own hardware, with my own money and on my own websites. The agents have real permissions: they place orders, sign on-chain transactions and edit live pages on their own.
The safety comes from limits enforced in code that the model cannot change. These notes cover how those limits are built, what the agents do within them and where things went wrong anyway.
What is running
Trading
Futures agent
Trades perpetual futures on its own schedule. Every order passes a risk gateway that enforces position size, leverage and loss limits, whatever the model decides.
On-chain
On-chain agent
Manages a small Solana wallet. A separate signer checks each transaction against a policy: allowed programs, caps per trade and per day, simulation before signing. The wallet never holds more than 2 SOL.
Operations
Website agent
Looks after seven WordPress sites on three hosts. Each change is backed up first, tested against the live page and rolled back automatically if the test fails.
Research
Research bench
Where both trading agents test ideas against historical data. The test plan is written before the run. Most ideas fail, and those results are published here too.
Coordination
Right hand
An agent that reads across all projects and keeps the threads together. It can read everything and change nothing. Actions go through the specialist agents.
Local
Local model
A 27B model on a single RTX 5090 handles the morning routine: mail triage, document filing, briefings. Cloud models take over where it falls short.
How I work
Limits live outside the model
Anything that must never happen is blocked in code, in a separate process with its own credentials. Prompts only describe what should happen.
Paper before live
A new strategy runs in a paper book first and goes live only after the numbers hold.
Failures get written up
A stop order that was rounded to zero. A faulty RAM module that silently corrupted data. A backtest that looked good until it turned out five trades carried all of it. These are the notes I learn the most from.
Latest notes
-
One agent, seven websites
An agent maintains my WordPress sites on its own. It never sees a hosting password, and every change is saved, tested on the live page and rolled back if the test fails.…
-
The stop order that rounded to zero
A rounding function turned a stop price into 0, my own monitor cancelled a working stop-loss, and a position sat unprotected for a night. What went wrong in plain code, and what…
-
Why the risk limits don’t live in the prompt
My futures agent decides on its own what to trade. How much it can lose is enforced by a gateway it cannot change. What the gateway checks, and why it refuses orders…
What this is not
Chainwhizz offers no consulting, coaching or investment advice, and nothing here is a recommendation to buy or sell anything. The trading agents are experiments with amounts I can afford to lose.
For my professional work on AI and digital assets in banking, see finwyz.