money agent safety harness

The problem

Developers gate autonomous agents on profitability while ignoring containment. When an agent is live, the first time you learn its "blast radius" is usually during an incident where a withdrawal-enabled key or a state mismatch drains the account.

What it does

• Enforces a 6-layer engineering gate to bound the financial risk of any agent that moves real funds.
• Verifies "capability floors" by querying venue permission endpoints with strict fail-closed logic.
• Audits paper-trading harnesses for a failure inventory rather than just profit/loss numbers.
• Identifies data-dependency traps where kill switches rely on the same corruptible state they are meant to guard.
• Validates state reconciliation loops to prevent losses from duplicate submissions or restart amnesia.

Frameworks & tools

Designed for any agentic execution stack. Includes specific verification patterns for venue APIs like Binance and libraries like CCXT.

Why this beats prompting it yourself

Most prompts fall into "safety theater" by accepting absent fields as safe. This skill specifically hunts for fail-open defects, such as .get(field, False), that human reviewers and standard LLMs habitually overlook. It transforms a vague "is this safe?" into a rigorous, evidence-based engineering audit.

Use cases

• Pre-live audit before funding a new trading bot or payment agent.
• Blast radius assessment for inherited or "vibe-coded" agents already running in production.
• Post-incident analysis to identify which containment layer failed and how to patch it.
• Creating a deployment gate that prevents bot boot-up if API permissions cannot be positively verified.

Known limitations

Not a judge of strategy or profitability. Does not replace legal or regulatory analysis for fund management. Some venues lack permission endpoints, requiring mandatory compensating controls.