The first time I tried to automate earnings-season research, I built a script. Feed it a ticker, it would fetch the transcript, run a sentiment pass, drop a summary into a spreadsheet. It worked, for about a week. Then a company changed its call format, added a Q&A session before the prepared remarks, and the whole pipeline silently produced garbage. Nobody noticed until an analyst asked why the "summary" was three lines about the operator's introduction.
That failure is the difference between automation and agentic work, and it's the reason I rebuilt what became StockLens around agents rather than scripts.
A script executes steps. An agent owns an outcome.
A script is a fixed sequence: do A, then B, then C. It has no model of what it's trying to achieve, only what it's supposed to do next. The moment reality diverges from the shape the script expects, it breaks, or worse, it doesn't break, it just produces a wrong answer with total confidence.
An agent is given a goal, "extract the guidance figures from this transcript and flag anything that contradicts last quarter's stated targets", and works out the steps itself. If the call has an unexpected structure, it adapts. If a number is ambiguous, it can decide to flag it rather than guess. That's not a philosophical distinction. It's the reason the Daily Updates and thesis-monitoring pieces of StockLens can run unattended across six markets every morning without someone checking that nothing broke overnight.
Why this matters more in banking than anywhere else
Most industries can tolerate a wrong automated output showing up quietly and getting caught later. Investment research cannot, not because the technology is different, but because the cost of an unnoticed error compounds through a decision that touches real capital.
This is where agentic systems have an advantage people don't expect: they're more auditable than the manual process they replace, not less. When I built the transcript intelligence engine, the requirement wasn't just "get the right answer," it was "show exactly which page of which document the answer came from." A human analyst skimming a transcript at 11pm doesn't leave that trail. An agent, built correctly, does it by default, because it has to cite its source to produce its output at all.
That's also why I've stayed deliberately skeptical of pure end-to-end model output for anything that touches a number a fund will act on. StockLens uses agents for orchestration, deciding what to fetch, what to check, what to flag, but the actual facts get pulled through retrieval with citations, not generated from a model's memory. The agent's job is to know what to look for and to keep working until it finds it, not to guess.
What actually breaks when you build these
The honest version of this story includes the failures. Early agent loops I built would occasionally get stuck re-checking the same filing, technically "working" while burning time without converging. The fix wasn't a smarter model, it was better scaffolding: explicit stopping conditions, a hard cap on retries, and a fallback to "flag for human review" instead of "keep trying forever." An agent that knows when to give up gracefully is more valuable than one that never gives up at all.
The other thing nobody tells you: agents are only as good as the boring infrastructure underneath them. Reliable document parsing, consistent citation formats, a data layer that doesn't silently drop a row, that's most of what makes an agentic workflow trustworthy. The agent reasoning is almost the easy part.
Where this goes next
The next stage isn't more autonomous agents, it's agents that know the boundary of their own competence. An agent that reads a 10-Q and flags "management changed the wording on this risk factor from last quarter" is doing something genuinely useful. An agent that decides, on its own, what that change means for your position is doing something a compliance officer should have opinions about before it ships.
That boundary, between agents that surface facts with full provenance and systems that make judgment calls, is exactly where I think the interesting and hard work is for the next few years. It's also why every workflow I build keeps a human decision at the end of the chain, not because the technology can't go further, but because in this industry, it shouldn't, not yet.