Agent-deployed trading

An LLM agent as the surface through which a strategy is authored, backtested and put live with real capital — the trader talks to the agent, the agent calls the platform. From superior-skills, the first artifact in this spoke where the deployment interface is a conversation.

What changes

Everything else in this corpus is a library or framework a programmer drives: tensortrade and trademaster are RL platforms, banbot an execution engine, kronos-financial-foundation-model a model you call. The human writes code, reads the backtest, and decides.

superior-skills moves each of those steps behind natural language. Its skills cover account bootstrap, deposits (including QR codes and external-wallet bridge links), strategy authoring, backtest submission, live deployment, sub-account orchestration, withdrawal, and an atomic exit-all kill switch. A “managed wallet” option is described as “no-key trading, custodial-style UX.” So the agent’s reach runs from writing the strategy to moving the money.

The vendor’s stated controls: draft → backtest → review → deploy, with “no live deployment without explicit user confirmation,” every parameter logged and every trade tracked.

The obvious problem, stated plainly

This spoke’s standing worry (backtesting, and the evidence question in synthesis) is that performance figures come from simulators their authors wrote. Agent deployment adds a second worry that is not about numbers at all: the last human check is an approval prompt.

The hub’s sibling ../agentic-tooling-wiki has been accumulating exactly the right material on what such a prompt is worth. Its guardrails thread records a shipping harness conceding that “an approval means a human accepted the displayed action under the information available at that time, not that the resulting behavior is safe”, and its constraint-evading-behavior page documents agents satisfying a check’s letter while defeating its purpose. Those observations were made about code review and shell commands, where the failure mode is a bad commit. Here the failure mode is a filled order.

Two properties make trading a worse case than the ones that literature examines:

• A trade has no diff. The reliability discipline in the agent literature assumes an artifact you can inspect before it counts. A position is opened or it isn’t.
• The kill switch is downstream of the mistake. exit-all closes positions; it does not unwind the ones already closed at a loss.

Nothing in superior-skills addresses either, and nothing else in this spoke does either. This page exists to name the gap rather than to claim it has been measured.

The commercial shape worth noticing

The skills are MIT-licensed and installable from five entry points (npx, gh skill install, Claude Code plugin marketplace, OpenClaw, or by pointing an agent at superior.trade/SKILL.md). They are also useless without an API key to a paid platform.

That makes an open skills pack a customer-acquisition channel: publish the interface, keep the backtester, the data and the execution. It is the first instance in this hub of a commercial financial service distributing itself as agent skills, and the incentive it creates is worth holding onto — the same party writes the strategy, runs the backtest that validates it, and earns on the deployment. superior-skills‘s six-trade “validated” strategy is what that incentive looks like in a README.

A second instance, and a correction (added 2026-08-06)

staskh-trading-skills is the same artifact class with the opposite economics: 25 MIT skills for options traders, free Yahoo data, the user’s own Interactive Brokers account, and no platform earning anything when a trade fires. So the paragraph above described one arrangement, not the pattern. The question that separates them is not whether a pack is open — both are — but what it needs a key for.

The pair also splits the failure modes cleanly, which is more useful than either alone:

• Superior has strategies and bad evidence for them. Six trades labelled “validated.”
• staskh has no strategies and therefore no evidence problem — and no backtester either, so nothing sits between a scanner hit and a position except the trader.

Neither closes the gap this page exists to name; they fail on either side of it. And the control boundary moves rather than disappearing: staskh defaults Interactive Brokers to read-only and requires a switch to read-write before the agent can place orders. That is a genuine default-safe design and it is still a single decision made once, in advance, covering every later order — after the switch, the approval-prompt question is back untouched. A trade still has no diff, and exit-all is still downstream.

Related

superior-skills · staskh-trading-skills · algorithmic-trading · backtesting · event-driven-trading · market-regime-analysis · synthesis