Taming AI agents - understanding is demonstrated, not declared

This is the third edition of The Perp Walk, and a catch-up: while the archive catches up to the present, each edition re-runs one of the strongest pieces from the past nine months. This one is from 11 July, on what it actually takes to make an AI agent reliable, and why the trick is to treat it less like a search engine and more like a junior analyst.

I have spent a lot of time recently trying to tame AI agents - not in polished demonstrations, but in the messy reality of getting them to deliver outcomes that are genuinely useful. One experiment sounded simple: scan public feeds across a sector, identify the lowest-cost, high-value options, assess the evidence behind them, and determine which ones were actually worth trusting. The theory was compelling, but the execution was humbling.

The agent could find information quickly, compare options, summarise claims and produce a polished-looking report, yet the first few versions were not good enough. What made the experience fascinating was what happened after each review cycle.

Every time we debriefed the output, the agent confidently explained that it understood the brief: it confirmed the objectives, acknowledged the feedback, and described how the next version would address the gaps. Then the next version arrived, and the same problems appeared, again and again. Every round of review, triangulation, debrief, and re-instruction followed the same pattern - a convincing explanation that the requirements were understood, followed by an output showing the understanding was incomplete.

The challenge was not intent; it was execution. It took multiple iterations to teach the agent what data actually mattered, because it needed guidance on which signals were meaningful, which sources deserved more weight, and how much information was required before drawing conclusions.

Then came the less glamorous problems. The output would occasionally end halfway through a sentence, tables would lose context, and important caveats would disappear. The analysis looked professional, but the underlying process was still fragile. The lesson is one that every leader working with AI agents needs to understand: an agent telling you it understands the task is not the same as an agent demonstrating it understands the task.

The breakthrough came when we stopped treating the agent like a search engine and started treating it like a junior analyst, and we designed the operating model around it:

Clear definitions of quality
Explicit evidence requirements
Source reliability rules
Completion checks
Structured review loops
Human judgement at critical points

AI agents are powerful, but capability alone does not create reliability. The organisations that gain the most value will not be those that simply deploy more agents; they will be the ones who understand the discipline required to turn probabilistic outputs into dependable business outcomes.

So the question every leader should ask before scaling an AI agent is this: have we taught the agent how to produce the answer, or have we only taught it how to tell us it understands the question?

First published on markneely.co: https://markneely.co/writing/7481696954494709760/