Work
What we have built and run so far. Every figure says where it comes from.
Product
PolySimulator
A simulation platform for event markets. Traders practise on live Polymarket markets with paper money; trading software runs against the same markets through an API that follows Polymarket's official client. Orders fill level by level against the real order book, with the venue's fees.
- Live since
- November 2025; first account 30 November 2025
- Accounts
- 7,000+, of which 3,700+ have placed a trade
- Simulated trades
- 390,000+ since launch
- API
- 80 accounts with an API key; Python SDK on PyPI since June 2026
PolySimulator production database, 16 October 2026. SDK release dates from PyPI.
How it is built
Engineering and operations run by agents
PolySimulator's code is written by AI coding agents and reviewed by a second AI model where the risk calls for it. Agents merge into our working branch and run staging themselves; in production they act under a standing permission from us. They also run the day-to-day operations.
- Changes merged
- 2,699 in 90 days, 980 of them in the last 30Merged pull requests. GitHub, PolySimulator's main repository, 16 October 2026.
- Availability
- API 99.98% since 9 August 2026, measured by two independent probes. Not an SLA; see the status page.PolySimulator status page, 16 October 2026.
- Lessons
- 188 written lessons in a shared memory of 1,686 notes, synced across machines every 20 minutesOur agent memory, 16 October 2026.
Watching
Agents watch our servers, the data capture and the public status of every service, and open a ticket when an alert fires.
Housekeeping
Backups, data reconciliation and clean-up jobs run on schedules the agents set up and check.
Briefings
Every morning and evening we get a briefing on what changed and what needs a decision from a person.
The system behind it
How the agents are organised
The setup that builds and runs PolySimulator, and the one we bring to clients.
Mistakes become lessons
When an agent gets something wrong, or one of us corrects it, the agent writes down what happened and what would have prevented it. The lesson reaches every other agent, and where it can, it becomes an automatic check instead of a reminder. The agents get more experienced as a team, without anyone retraining them.
Standing agent sessions
Long-running agent sessions on two always-on machines, each in a fixed role: operations, business, a research engagement, this website. They message each other directly instead of rebuilding what another one owns.
A shared memory
One versioned memory store per project, synced across machines every 20 minutes, so every session starts with the same context. Facts carry a lease and are checked again when it runs out; nothing is deleted silently, so a correction is never lost.
Tickets and a board
Every request becomes a ticket with a status, and a board shows which session is doing what. Each session starts from the same operating manual and the open tickets.
Orchestrated teams
For larger work, an orchestrating agent plans the phases, sends out a team of agents in parallel and has a critic check what comes back before anything is merged or sent.
Data
Our market data
Since 30 July 2026 we record Polymarket's full order-book feed ourselves: on the order of 100 million updates an hour, tens of terabytes so far, with settlement reference prices alongside. It is the ground truth for our research on fill fidelity and point-in-time replay, and the heart of what we build.
The update rate is an order of magnitude, measured in August 2026; the archive size is from October 2026.
Our researchOpen source
Public code
polysimulator-sdk
The official Python SDK for PolySimulator's paper-trading API, with a drop-in client for code written against Polymarket.
polysimulator-docs
The API documentation, synced from the main repository and readable by AI coding tools.
polysim-bootstrap
The one-command setup for our development machines: toolchain, AI coding agents and their tool servers, with secrets pulled from a vault and none kept in the repository.
super-board, extended
A build, test and review loop for coding agents, driven by a project board. It is Eric Tech's open-source tool; Wladimir extended it with builder and tester agents on models from other providers, a binding review by a second model family, and ideas broken down into detailed tasks before any agent starts.
Research
Research engagements
We take on research engagements on prediction-market infrastructure: resolution, oracle economics, execution and capacity.
Ask about an engagementWant this for your company?
Write to [email protected]. We reply in German or English.