When you want an agent to do a real business task, three problems can get in the way. It may fail to follow rules across a long conversation. It may need detailed instructions. And you may not know what to ask for. A narrow agent built for one job is designed to deal with all three. That design goal is not proof that a particular product succeeds; the vendor still needs to show its tests.
Check 1: Does it keep your rules over many runs?
Research on long conversations is a warning here. The ICLR 2026 paper LLMs Get Lost In Multi-Turn Conversation analyzed over 200,000 simulated conversations and reported an average 39% performance drop across six generation tasks in multi-turn versus fully specified single-turn settings. It tested fragmented, underspecified instructions, not every long conversation. The ACL Findings 2026 paper What Prompts Don't Say found underspecified prompts were twice as likely to regress across model or prompt changes. It also found that specifying every requirement does not consistently help, because requirements can conflict and models have limited instruction-following ability. These are lab tests, not tests of agents with a saved rulebook. So ask the vendor how they test that rules still hold after many runs.
Check 2: Are the long instructions already written?
Good instructions give the goal, the steps, the limits and the quality bar. The exploratory preprint DETAIL Matters, using 30 novel reasoning tasks and GPT-4 and O3-mini, reported better accuracy with more specific prompts, especially for procedural tasks. Its small task set and model-based correctness assessment limit the conclusion. Ask whether the instructions are built in and tested, or whether you must write them yourself.
Check 3: Does it start from one clear job?
A blank prompt box leaves the job, the instructions and the quality bar to you. The Carnegie Mellon study Why Johnny Can't Use Agents reviewed 102 marketed agents and studied 31 people using Operator and Manus. It found five usability barriers, including mismatched mental models, premature trust and weak recovery or collaboration. That is evidence about those tested tools and tasks, not a comparison proving vertical agents are better. A narrow agent starts from a defined job, such as intake or reply handling.
What to still be careful about
A forecast that vertical agents will boom is not proof of better performance or a lasting business. A thin wrapper around a big model may not last. Work in legal, health and finance still needs a person to sign off.
A short buyer checklist
- How do you test that rules hold across runs?
- Which instructions are built in?
- What is the one job it does?
- Where does a person approve?
Watch the companion video
Will Vertical AI Agents Boom? What the Research Does and Doesn't Say on Predicting Future. The video presents the thesis; the study limits and buyer checks above remain important.