Predictika · Agent trade gate

An AI trading agent's orders, checked before they are placed

An AI trading agent proposes orders for a managed brokerage account, whether a person asked in chat or a scheduled Loop runs them: stocks, ETFs and crypto, up to three orders at a time. Before they are placed, the agent sends them to Predictika, which checks the orders against the account's mandate and limits, and returns a verdict, every reason, and the figures it computed, such as position size and cash left, so the agent's own claims can be checked. Designed for Robinhood's agent trading, and usable by any brokerage, exchange or firm that runs trading agents: Loops, optional trade approvals and state crypto rules reflect features of Robinhood Agents. Kestrel Capital is a fictional company, and these are not Robinhood's rules.

Scope: Buy and sell orders for stocks, ETFs and crypto that an AI agent proposes for a managed brokerage account, from chat or a scheduled Loop run, in plans of up to three orders. See everything this validator covers

It handles any order involving those instruments, not only the scenarios shown on this page. It was designed around Robinhood's agent trading, and the same checks apply to any company running trading agents, whether the agents are its own or its customers'.

  • Limit prices — Compares the limit with the reference price and blocks anything more than 3 percent away.
  • Concentration — Computes the position after the order from the current holding and blocks above 25 percent of the account.
  • Stated figures — Computes all three and blocks when the agent's figures differ.
  • Quantity slips — Prices the order itself, blocks it above the single-order limit, and blocks when the stated total differs from the computed value.
  • Loops — Keeps the day's running total against a daily budget across runs.
  • Mandate — Checks the instrument's class against what the account is allowed to hold.
  • Approvals — Holds orders above the sign-off line, and crypto orders where a state requires approval, for a person.

Not covered: options, margin, short selling and futures. The validator does not place orders itself: it returns allow, needs approval or block with every finding.

The AI trading agentBuilds the orders from the chat or the Loop schedule and says they are ready.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.

9 scenarios in 5 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

The AI trading agent says

Illustrative agent statement. The validator receives the structured request below.

Proposing agent: AI trading agent (scripted)

Before sending, the agent must call Recorded run

Predictika validator


      

        
        
        
      

      
Waiting for the call.

      
    

  

  
  

    
Where an AI agent gets this wrong on its own
Click to expand
    
Our assessment of what an AI trading agent does when nothing checks its orders. It is a judgment about tendencies, not a claim that an agent can never get these right.

    
Failure modeAn AI agent on its ownPredictika
Limit pricesError-prone
Types or copies a limit price far from the market and sends it.		Compares the limit with the reference price and blocks anything more than 3 percent away.
ConcentrationError-prone
Keeps adding to one holding until it dominates the account.		Computes the position after the order from the current holding and blocks above 25 percent of the account.
Stated figuresError-prone
Reports a position, a cash balance or an order total from memory or a rough estimate.		Computes all three and blocks when the agent's figures differ.
Quantity slipsError-prone
Enters a quantity with an extra digit or a decimal slip and describes the order as it meant it.		Prices the order itself, blocks it above the single-order limit, and blocks when the stated total differs from the computed value.
Loops that overrun a daily budget or order rateError-prone
Runs on a schedule with only its own rule of thumb, so each run looks fine alone.		Keeps the day's running total against a daily budget across runs.
MandateError-prone
Picks an instrument because it fits the idea, not because the account may hold it.		Checks the instrument's class against what the account is allowed to hold.
ApprovalsError-prone
Treats a human check as something it can skip when it is confident.		Holds orders above the sign-off line, and crypto orders where a state requires approval, for a person.

  

  

    
Public incidents two scenarios resemble
Click to expand
    
The accounts, instruments and amounts on this page are invented. Two scenarios are inspired by publicly reported incidents.

    
ScenarioPublicly reported incident
A crypto order far larger than the agent statesInspired by a publicly reported case from February 2026, in which an AI crypto agent sent about $442,000 of tokens instead of about $310, likely a decimal error. The coin, account and amounts here are invented.
Three very large sell orders with a wrong stated totalInspired by a publicly reported case from May 2022, in which a bank trader entered a $444 billion order instead of $58 million and reportedly dismissed hundreds of warnings. The account, instruments and amounts here are invented.

  

  

    
What this validator catches: its checks
Click to expand
    
Every check runs on every order in a plan, in any order the choices arrive; one problem never hides another. Blocking findings stop the request; approval findings need a person to sign off; everything else is computed.

    
1Instrument class not allowed by the account mandate
2Order above the single-order limit ($200,000)
3Limit price more than 3 percent from the reference price
4Holding above 25 percent of the account after the order
5More than 4 orders in 10 minutes
6Buys exceed cash on hand (sale proceeds do not count until they settle)
7Scheduled Loop run over its daily budget
8Agent's stated position share differs from the validator's computation
9Agent's stated cash left differs from the validator's computation
10Agent's stated total order value differs from the validator's computation
11Crypto in a state that requires approval every time (approval)
12Order above $100,000 (approval)

  

  

    
All 22 test cases run on the live validator
Click to expand
    
GroupCaseVerdictOn this page
Fat-finger limit priceRoutine buy: 1,050 HLX, limit 80.50, 18.5 percent of the account afterAllowshown
Fat-finger limit priceFat-finger limit: 1,050 HLX with a limit of 90.00 against a reference of 80.00Blockshown
ConcentrationConcentration: 8,000 PNR takes the position to 30.8 percent and the order is $340,000Block
ConcentrationSmaller order: 3,400 PNR, 22.7 percent after, $144,500 needs sign-offNeeds approval
MandateMandate: 1,000 LVX3, a leveraged ETF, in an account that does not allow leveraged productsBlockshown
MandateIndex ETF instead: 540 SPRX, allowed instrumentAllow
State crypto ruleNY account, crypto: 20,000 VLT ($96,000) needs approval under the state ruleNeeds approval
State crypto ruleTX account, same 20,000 VLT order: no state ruleAllow
Loop budgetLoop run: 400 SPRX ($48,000) after $95,000 already spent today against a $120,000 daily budgetBlock
Loop budgetLoop run, smaller: 200 SPRX ($24,000), $119,000 for the dayAllow
CashNot enough cash: 1,100 HLX ($88,000) with $60,000 on handBlock
Stated figuresClaim mismatch on cash: the agent states $200,000 left, the validator computes $316,000Block
VelocityOrder velocity: 4 orders already in the last 10 minutes, one more would be the fifthBlock
Multi-step planClean rebalance: sell 1,000 PNR, buy 400 SPRX, both within policyAllow
Multi-step planEverything wrong at once: Loop run, 8,000 LVX3 at limit 90.00, little cash, 4 orders in the windowBlock
BoundariesBoundary: every limit exactly at its edge (25.0 percent, $100,000, 3.0 percent, 4th order)Allow
Rebalancing basketSale-funded rebalance: sell 2,300 PNR, buy 1,250 HLX and 800 SPRX with $150,000 of cashBlockshown
Rebalancing basketRebalance resized: sell 2,300 PNR, buy 900 HLX and 600 SPRX within cash on handAllowshown
Loop basket, New YorkNY Loop basket: 15,000 VLT, 300 AXN at limit 322.00, 700 SPRX, with 2 orders already in the windowBlockshown
Loop basket, New YorkNY Loop basket, trimmed: 10,000 VLT and 160 AXN at limit 311.00, two orders in the windowNeeds approvalshown
Inspired by real incidentsInspired by a reported AI crypto agent error: 62,000 VLT ($297,600) instead of 62 VLT, with the agent stating about $300Blockshown
Inspired by real incidentsInspired by a reported fat-finger rebalance: three sells each entered with an extra digit, $2.13 million instead of $213,000Blockshown

  

  

    
Scenarios and results
Click to expand
    
The recorded result of the validator for each scenario on this page: the request the AI agent sent, the verdict, and every finding, in the order they appear above.

    
1. Loop basket, New York: A Loop run of three orders that breaks three limits

    
The agent's request. This is the scheduled Loop run. Buy 15,000 Volt (VLT, a crypto coin) at the market, buy 300 Axion (AXN, a crypto coin) at a limit of $322.00, and buy 700 shares of Sprout 500 Index ETF (SPRX) at the market. The largest position will be 11.8 percent of the account and cash on hand after the plan will be $151,000.

    
Verdict: Blocked.

    
    
      
  • Limit price too far from the market. The limit price on order 2 is 3.87 percent away from the reference price of $310; the most allowed is 3.0 percent.
    • 
      
  • Too many orders in 10 minutes. This would be 5 orders in 10 minutes (2 already placed plus 3 in this request); the limit is 4.
    • 
      
  • Loop over its daily budget. $20,000 already spent today plus $249,000 in this run is over the $120,000 daily budget for Loop runs.
    • 
      
  • State rule: crypto needs approval. This request includes crypto and NY is a state where crypto orders need a human approval each time.
    • 
    

    
2. Loop basket, New York: The agent resubmits after fixing the errors

    
The agent's request. This is the scheduled Loop run. Buy 10,000 Volt (VLT, a crypto coin) at the market and buy 160 Axion (AXN, a crypto coin) at a limit of $311.00. The largest position will be 7.0 percent of the account and cash on hand after the plan will be $302,400.

    
Verdict: Needs approval.

    
    
      
  • State rule: crypto needs approval. This request includes crypto and NY is a state where crypto orders need a human approval each time.
    • 
    

    
3. Limit price: A limit price too far from the market

    
The agent's request. Buy 1,050 shares of Helix Systems (HLX, a stock) at a limit of $90.00. The position will be 18.5 percent of the account and cash on hand after the order will be $316,000.

    
Verdict: Blocked.

    
    
      
  • Limit price too far from the market. The limit price on order 1 is 12.5 percent away from the reference price of $80; the most allowed is 3.0 percent.
    • 
    

    
4. Limit price: The agent resubmits after fixing the errors

    
The agent's request. Buy 1,050 shares of Helix Systems (HLX, a stock) at a limit of $80.50. The position will be 18.5 percent of the account and cash on hand after the order will be $316,000.

    
Verdict: Allowed.

    
    
      
  • Within policy. No finding: the orders are within the account policy.
    • 
    

    
5. Rebalancing basket: A rebalancing basket with a position too large and not enough cash

    
The agent's request. Sell 2,300 shares of Pinecrest Retail (PNR, a stock) at the market, then buy 1,250 shares of Helix Systems (HLX, a stock) and 800 shares of Sprout 500 Index ETF (SPRX), both at the market. The largest position will be 25.8 percent of the account and cash on hand after the plan will be $51,750.

    
Verdict: Blocked.

    
    
      
  • Too much in one holding. Order 2 would make Helix Systems 25.83 percent of the account after the order; the limit is 25.0 percent.
    • 
      
  • Not enough cash. The buys total $196,000 but only $150,000 of cash is available. Sale proceeds do not count until they settle.
    • 
      
  • Agent's stated cash is wrong. The agent stated $51,750 of cash left after the orders; the validator computes -$46,000 (tolerance 100.0).
    • 
    

    
6. Rebalancing basket: The agent resubmits after fixing the errors

    
The agent's request. Sell 2,300 shares of Pinecrest Retail (PNR, a stock) at the market, then buy 900 shares of Helix Systems (HLX, a stock) and 600 shares of Sprout 500 Index ETF (SPRX), both at the market. The largest position will be 24.7 percent of the account and cash on hand after the plan will be $6,000.

    
Verdict: Allowed.

    
    
      
  • Within policy. No finding: the orders are within the account policy.
    • 
    

    
7. Mandate: A leveraged ETF in an account that does not allow it

    
The agent's request. Buy 1,000 shares of Tech Surge 3x Leveraged ETF (LVX3) at the market.

    
Verdict: Blocked.

    
    
      
  • Instrument not allowed for this account. Order 1 is Tech Surge 3x Leveraged ETF (levetf), a class this account is not allowed to trade.
    • 
    

    
8. Inspired by real incidents: A crypto order far larger than the agent states

    
The agent's request. Buy 62,000 Volt (VLT, a crypto coin) at the market. The order total will be about $300.

    
Verdict: Blocked.

    
    
      
  • Order too large. Order 1 is $297,600, above the $200,000 single-order limit.
    • 
      
  • Agent's stated order total is wrong. The agent stated orders totaling $300; the validator computes $297,600 (tolerance 100.0).
    • 
      
  • Large order: human approval. The largest order is $297,600, above the $100,000 limit. A human must approve it before it is placed.
    • 
    

    
9. Inspired by real incidents: Three very large sell orders with a wrong stated total

    
The agent's request. Sell 10,000 shares of Helix Systems (HLX, a stock), sell 20,000 shares of Pinecrest Retail (PNR, a stock) and sell 4,000 shares of Sprout 500 Index ETF (SPRX), all at the market. The orders total $213,000.

    
Verdict: Blocked.

    
    
      
  • Order too large. Order 1 is $800,000, above the $200,000 single-order limit.
    • 
      
  • Order too large. Order 2 is $850,000, above the $200,000 single-order limit.
    • 
      
  • Order too large. Order 3 is $480,000, above the $200,000 single-order limit.
    • 
      
  • Agent's stated order total is wrong. The agent stated orders totaling $213,000; the validator computes $2,130,000 (tolerance 100.0).
    • 
      
  • Large order: human approval. The largest order is $850,000, above the $100,000 limit. A human must approve it before it is placed.
    • 
    

  

  

    
Narration transcript
Click to expand
    
The text of the narrated walkthrough, step by step.

    
    
      
  1. Welcome to a demo of Predictika's Validator. An AI trading agent proposes orders for a managed brokerage account, whether a person asked in chat or a scheduled Loop runs them: stocks, ETFs and crypto, up to three orders at a time. Before they are placed, the agent sends them to Predictika, which checks the orders against the account's mandate and limits, and returns a verdict, every reason, and the figures it computed, such as position size and cash left, so the agent's own claims can be checked. Designed for Robinhood's agent trading, and usable by any brokerage, exchange or firm that runs trading agents: Loops, optional trade approvals and state crypto rules reflect features of Robinhood Agents. Kestrel Capital is a fictional company, and these are not Robinhood's rules.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. This is the scheduled Loop run. Buy 15,000 Volt (VLT, a crypto coin) at the market, buy 300 Axion (AXN, a crypto coin) at a limit of $322.00. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with three findings and one approval, all in one answer. The limit price on order 2 is 3.87 percent away from the reference price of $310; the most allowed is 3.0 percent. This would be 5 orders in 10 minutes (2 already placed plus 3 in this request); the limit is 4. $20,000 already spent today plus $249,000 in this run is over the $120,000 daily budget for Loop runs.
    6. 
      
  6. A scheduled Loop run in the New York account sends three orders. The Axion limit of $322.00 is 3.87 percent above its $310.00 reference, over the 3 percent limit. Two orders were placed in the last 10 minutes, so this would be the fifth, one over the limit of four. The run would also take the day to $269,000 against a $120,000 daily budget. The agent's stated position and cash figures are correct, so those checks pass: the blocks come from the plan itself. The crypto orders would also need approval in New York.
    7. 
      
  7. What this shows: for loops that overrun a daily budget or order rate, an AI agent on its own is error-prone for this.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the first order to Buy 10,000 VLT (Volt, crypto coin) at the market; current holding $120,000 and the second order to Buy 160 AXN (Axion, crypto coin) at a limit of $311.00; current holding $0. It removes the third order.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. Not blocked, but held for a person.
    12. 
      
  12. The agent trims the run to two crypto orders totaling $97,600, so the day reaches $117,600 and the window is exactly four, and it brings the Axion limit to $311.00, 0.32 percent from the reference. Nothing is blocked. Because the account is in New York and both orders are crypto, the validator holds the plan for a person to approve instead of releasing it.
    13. 
      
  13. Four problems caught in one call, fixed in one pass, and the corrected request now goes to a person for approval. That is the value of Predictika's validator: catching mistakes before the action runs.
    14.