
TreasuryForge Makes Agent Trades Approval-First
TreasuryForge is an autonomous treasury demo for simulated cash, crypto, and NSE equities, built on TrueForge for the WeMakeDevs × TrueFoundry Agent Harness Hackathon. Its loop computes risk server-side, stress-tests breaches in a sandbox, and pauses every trade for explicit human approval.
The achievement is not autonomous trading; it is making the approval boundary executable. TreasuryForge also shows why agent demos need adversarial review, since early versions contained both an approval bypass and a retry-driven double-execution bug.
- –MCP keeps mutations behind a tool boundary, while server-side checks prevent model-supplied reasoning from becoming risk authority.
- –Runtime approval checkpoints are materially stronger than “ask first” instructions embedded in prompts.
- –Qodo exposed a direct MCP bypass caused by an unauthenticated reset endpoint and overly broad network binding.
- –The project remains demo-grade: local sandboxing is Linux/WSL2-only, backtesting lacks real historical prices, and the frontend has no automated test suite.
- –The pattern generalizes well: deterministic checks and human authorization belong at the mutation boundary, not inside the model’s context.
DISCOVERED
1h ago
2026-08-30
PUBLISHED
3h ago
2026-08-30
RELEVANCE
AUTHOR
kinmorychin