Welcome to the first weekly roundup from the Agentic Trading Use Cases hub. Each week I collect a handful of real, dated items about AI agents, broker connections, testing and review. I only include links I opened and read myself.
This is education about process. It is not a list of trade ideas, and nothing here says any setup makes money.
This week's theme: when an AI agent can reach a brokerage account, who holds the keys, and how do you know your tests match what actually runs?
Approval gates: who holds the keys?
Four items on how much authority an outside agent should get, and how brokers and tool makers are drawing the lines.
Schwab and JPMorgan keep outside agents away from moving money
Oct 16, 2026 · The Next Web, reporting a Bloomberg TV interview
Schwab's CEO told Bloomberg Television that Schwab is blocking outside AI agents that try to move client money. The Next Web reports these agents are stopped tens of times a week. Instead, Schwab plans controlled MCP servers that let outside AI tools check balances and do research, but not move money. The same report says JPMorgan's CEO took a similar line on an earnings call earlier this week.
Axi connects MetaTrader 5 accounts to Claude, with approval in chat
Oct 14, 2026 · Finance Magnates
Axi said eligible clients can connect MetaTrader 5 accounts to Claude through MCP, in alpha testing. Per the connector's website, any trade or order action needs the client's explicit approval in the chat. The article notes one exception: a client can set a specific action to “Always allow” in Claude's permission settings. It also notes that Match-Trade opened its MCP to retail traders on Oct 5, with an option to require approval for each action.
Robinhood Agents: approvals on by default, a dedicated account, and customer-owned risk
Unveiled Sept 29, 2026 · Robinhood Help Center; PYMNTS, Oct 1
Robinhood describes agents built into its app that trade in a dedicated Agentic account, and you choose how much to fund it. Trade approvals are on by default, so the agent proposes and you place the order. With approvals off, it can place eligible trades without confirming each one. Robinhood's disclosures say customers assume all risk and that Robinhood does not control, supervise, monitor or audit agents.
tastytrade's MCP server: dry run first, then a one-time confirmation token
Developer docs, checked Oct 16, 2026 (docs page is undated)
tastytrade's open-source, self-hosted MCP server will not submit or change an order until it gets a single-use confirmation token from that order's own dry run. Per the docs, the token is valid for 60 seconds and is tied to a hash of the exact arguments, so you cannot preview one share and submit a thousand. There is also a read-only switch and a configurable cap on an order's buying-power impact. The docs warn that the default endpoint is production unless you choose the sandbox.
Backtest versus live: where the drift hides
Two items on a problem every builder meets: the thing you tested is not always the thing that runs.
Where Backtests and Live Trading Part Ways: Six Defects from One Intraday System
Oct 15, 2026 · arXiv 2610.21846 (working paper)
Two independent researchers compared live days of one intraday US equity system with a backtest replay of the same days and found six defects. Three were data problems: a default feed carrying a small fraction of real opening volume, prices adjusted for dividends that live orders never see, and data the live system downloaded but the backtest did not. Two were code problems: a sizing step that exceeded a live account limit, and UTC timestamps read as New York time. One was configuration that never reached the order code. They close with a ten-point checklist and make no claim about profitability.
MintEval: does the code do what you asked for?
October 2026 · arXiv 2610.03080 (early-stage v0 benchmark)
MintEval targets a quiet failure: an AI writes strategy code from a plain-language description, the code runs and a backtest plots, yet it does not behave as described. The authors run generated and reference code bar by bar and compare actions, not profit. In their small v0 test on BTCUSDT 15-minute data, a frontier model on a 200-task subset still failed silently on 27.5% of tasks. A code-reading LLM judge passed every run it was given, including those silent failures.
Journaling and review: tools keep shipping, so check the numbers
One item on the tools many of us use to review our own trading.
TradeZella weekly release notes: timezone, look-ahead and import fixes
Oct 13 and Oct 6, 2026 · TradeZella changelog
TradeZella, a trade journaling and backtesting platform, publishes weekly release notes. The Oct 13 notes list fixes including backtests set to a US/Eastern session that ran on UTC time, a look-ahead bug where a divergence signal triggered before its swing was confirmed, and an IBKR sync that imported cash-settled index options with a $0 close price and wrong P&L. The Oct 6 notes announce a beta MCP server for its AI assistant. This is a vendor changelog, not an endorsement.
Takeaway and a short process checklist
Process checklist for agent-connected trading
✓ Separate read access from order authority.
✓ Preview the exact order, then approve exactly that order.
✓ Know every setting that can switch approvals off, and who can change it.
✓ Replay real live days through your test harness and compare decisions, not just totals.
✓ Check data-feed identity and clock handling (time zones, daylight saving) in every request.
Education only. This is not financial advice or a recommendation to buy or sell anything. Items are plain-language summaries of public sources, linked for reference and not as endorsements. Check details at the source. Browse more on the Agentic Trading Use Cases hub.
— helixtrader36
Comments
Post a Comment