Approach
Velytra treats an AI trading desk as an engineering system. Agents may be autonomous, but every order path is gated, every rule lives in one place, and every outcome is measured against a benchmark that could prove the strategy wrong.
Map the book and the mandate
We start with your accounts, brokers, and the goals the system is graded on. A mandate written in plain English keeps the agents and the humans on the same page.
- Brokerage inventory: which accounts, which APIs, which approvals already exist.
- Goal hierarchy with explicit tolerance bands and what bends when goals conflict.
- Current workflow audit: what is manual, what is daily, what breaks silently.
- Tax posture and Trader Tax Status review where relevant.
Connect brokers behind one interface
Strategy code depends on a single broker abstraction. Adapters for Schwab, Tastytrade, Alpaca, and SnapTrade plug in underneath and must pass the same contract tests.
- MCP servers with OAuth, token refresh, and account identifier scrubbing.
- Configuration, secrets, and runtime state kept strictly separate from code.
- Every upstream tool classified as allowed, blocked, or removed; unknown tools denied.
- Paper venue first, with the real-money path unwired until you decide otherwise.
Autonomy inside hard limits
Agents screen, propose, and report on their own. Orders pass through risk checks in code, then through a human approval, before anything reaches a broker.
- Circuit breakers for daily loss and leverage, consulted before any order.
- Default-deny order guards with per-order notional caps and session-hours checks.
- Slack approve/reject buttons, then deterministic execution with no model in the loop.
- A kill switch for every scheduled agent and a watchdog that restarts stale jobs.
- Changes to hard rules require owner authorization, in code review and in the playbook.
Grade the strategy, not the activity
A strategy that cannot be falsified is a hunch. We replay rules on years of option history, reconcile live results against the replay, and let the evidence edit the playbook.
- Point-in-time universes and mid-price fills so the backtest cannot flatter itself.
- Post-mortem on every closed trade with rule-level attribution.
- Honest peers: PutWrite and buy-write indices, income ETFs, and the index itself.
- Weekly reviewer rewrite of the playbook, with anti-drift checks against the mandate.
- Numbers never travel without their assumptions and the condition that would flip them.
How the system stays trustworthy
One source of truth
Shared logic lives in one core library. Agents, dashboards, and skills import it rather than reimplementing a rule.
Same standard, both seats
The protocol the paper agent follows is the same one used for real accounts. Paper is the venue, never a discount on care.
A quiet bot is a bug
No-trade droughts and silent job failures are defects. Ops channel reviews run daily and fix what the alerts expose.