Research
Why asking AI trading agents beats telling them what to do
The most useful practical result from a year of agent research is boring: instruction quality dominates raw model capability. For finance teams, the prompt is now a first-class engineering artefact.
The FinTech Global piece surfaces something practitioners already knew but rarely wrote down: an agent asked to find the strongest risk factor in this earnings release given recent macro context outperforms an agent told to look for revenue miss, guidance cut, or margin compression. The instruction that opens the search space produces better analysis than the instruction that constrains it, provided the agent has the underlying capability. That inversion is not intuitive. Most finance teams still write agent prompts like they are writing a runbook for a junior analyst.
For a team running a live agent stack, the operational implication is that the prompt is code and belongs in version control. Not a document. Not a shared notes page. Code, reviewed on the same cadence as any other change to the strategy, with regression tests attached. The equivalent artefact in a rules-based quant shop is the signal definition. Nobody would edit a signal definition in a document; the same discipline should apply to the instruction that steers an LLM-based analyst. Teams that treat prompts as prose and models as code end up shipping regressions they cannot trace.
There is a subtler point behind the finding that matters for governance. If asking outperforms telling on complex tasks, then the specific asking pattern becomes part of the alpha. A team's prompt library encodes real knowledge about how to elicit useful analysis from the underlying model. That library is now an intellectual-property asset that should be handled as such: access-controlled, backed up, and treated as leaving with the wrong departing employee. Very few AI-driven finance teams have this operational posture yet. The ones that build it early will save themselves an unpleasant conversation later.
For the regulator-facing surface, the finding also raises a documentation question. FCA MAR 7A and equivalents in other jurisdictions require documented change control on algorithmic strategies. If the prompt materially changes strategy behaviour, and small prompt changes can produce large behavioural shifts on complex asks, then the prompt is an in-scope algorithmic component for change-control purposes. That is not the current market practice. It should be. The teams that get audited first will discover this the expensive way; the teams that write it into their SOPs now will not.
Read the original: FinTech Global - Why asking AI trading agents beats telling them what to do. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.