Agentic trading: when agents place the orders
When agents place orders, the failure that costs money is a valid-looking order produced by a broken chain of reasoning.
- Sector
- Trading platforms and regulated fintech
- Running agents like these?
- Talk to us
- Signal
- read
- Risk check
- passed
- Order
- ready
Exposure
Syntrox verdictBlock
Breaches the exposure ceiling. Blocked before execution and escalated to an operator.
Other customers’ agentsstill trading
The setting
The platform in this case study is betting that trading stops being something a person does at a screen and becomes something an agent does continuously on their behalf: watching the market, analysing conditions, sizing positions and executing.
Under the hood that is several agents working together. A manager agent breaks the trader's intent into parts. An analysis agent reads the market. A risk agent checks the position. An execution agent places the order. From the day it goes live, software moves customer money without a person approving each decision.
What goes wrong
The failure everyone expects is a chatbot saying something wrong. That is not the failure that costs money here.
The failure that costs money is a consequential action produced by a defective chain of reasoning. The analysis agent misreads a signal. The risk check passes because its input arrived in the right shape, not because it was right. The order goes out. Afterwards every line in the log looks fine, and the money is gone anyway.
Three things follow:
- The orchestration is a black box. When agents hand work to each other, nobody can see what the system as a whole decided.
- A decision cannot be reconstructed. Not just what the agents did, but why, given what they knew at the time.
- There is no record of authority. Nothing shows which agent acted, under whose authority and within which limit.
Why the obvious tools don't close it
- Observability records what happened after settlement. Useful for debugging last week; no help this morning.
- Prompt guardrails filter text going in and out of a model. They have no view of an order. A polite, well-formatted instruction to place a bad trade passes every text filter.
- Pre-launch evaluations run against a fixed test set. The failures that matter happen at runtime, against market states nobody wrote a test for.
Where regulators have written rules for trading algorithms, they ask for the same things: an identifier on every order, a kill switch and an audit trail. SEBI's framework for Indian brokers, binding since 1 April 2026, is one example.
What Syntrox does
Syntrox sits between the agents and the systems they can act on, independent of the agent framework.
- Before an order executes, the action is checked against the platform's own rules: position limits, exposure ceilings, permitted instruments, and the conditions under which an agent may act without a person. Each verdict names the rule it applied.
- When an action breaches policy, it is blocked or escalated, and the agent responsible can be isolated without halting trading for every other customer. Isolation triggers automatically on per-agent spend ceilings; for other breaches, an operator isolates the agent with one action.
- Afterwards there is a record: what the agent attempted, what it knew, the verdict, the rule and what a person did next. It exists whether or not anything went wrong.
- Everything runs inside the platform's own environment. No trading data, positions or customer information leaves.
Two numbers, always labelled: hard rules run inline in under 20 ms; a judge verdict from Nucleus, our own 4-billion-parameter model, takes about 145 ms.
We built agents we couldn't ship, because we couldn't explain them to ourselves — let alone to a regulator. Syntrox is the layer that made autonomous trading something we could put our name on.
How a pilot is measured
A 30-day shadow test, where Syntrox observes and judges but blocks nothing:
| What we measure | Why it matters |
|---|---|
| Unsafe actions caught before execution, with the rule cited | Direct loss avoided |
| False-positive rate | The real cost of the control, reported as measured |
| Latency added | It has to be invisible at trading speed |
| Share of consequential actions with a reconstructable reasoning chain | What an auditor or exchange asks for |
| Where failures start: inputs, reasoning, retrieval, tool choice or goal-seeking | Tells engineering what to fix |
| Time to answer "why did the agent do that?" | Today hours of log reading; the target is one query |