10 paper-trading agents show parallel calls can bypass spend limits — where should the hard cap live?
Accomplished_Fun_408 · reddit · 2026-09-29
A RockFlow engineer observed 10 AI agents trading $1M of fake money each via MCP and distilled a design question: where should spend/position limits live?
- One agent put 84% of its account into a next-day call option (and won); another went 100% single-contract and lost half the account in three sessions. Neither had any sizing rule — the only check they hit was buying power.
- Agents firing parallel tool calls can pass the same "remaining allowance" check repeatedly unless the reservation and check are atomic. No prompt-level policy fixes this.
- Paper fills at screen price with no depth made every mistake look cheaper than with real money.
Limits can live in three places: the prompt (cheap, ignored when the model is confident), the tool schema (visible but unenforced), or the account/an order gateway (deterministic but must be built and every order path routed through it). The team is moving to the third.
More from coding & agent
- Dev builds multiplayer browser game with Opus in just 3 hours — mhmazur · 2026-09-29
- Autonomous AP Auditing: Building a Precedent-Governed Copilot with Stateful Memory — ashokyuva · 2026-09-29
- A vibe-coded website and product just landed a $750K seed round — gaganghotra_ · 2026-09-29
- Turning a Reviewed Shot List into Multishot Video Prompts with a VL Model and a Field Template — kim_deadja4951 · 2026-09-29
- Building an Agent Harness with Jev: 15 Tests Reveal Which Decision Points Actually Help — Prompt Engineering · 2026-09-29
- Undoing one AI edit without losing the rest: change-record replay with staleness checks — memokris · 2026-09-29