The Three Rules for Managing AI Agents Before They Manage Your Budget

The team behind the Exponential View newsletter, who between them push more AI-agent work through their business in a week than most Malaysian SMEs will attempt this year, recently sat down to ask what they'd got wrong about managing agents over the past six months, and the answer is more useful to a ten-person company than to them, because a large team can absorb a wasted afternoon of agent work in a way a small one cannot.

Azeem Azhar and his colleagues published the result as seven lessons for managing AI agents, an update to a similar piece they wrote in April 2025, before agents could be trusted to work unsupervised for anything longer than a few minutes. That has changed quickly: OpenAI's own figures, cited in the piece, show that by May this year roughly a quarter of Codex users were making at least one request a month for work that would take a person a full eight-hour day to finish, up from just two percent the previous December. Three of the seven lessons, the ones freely available before the article goes behind Exponential View's paywall, are worth a Malaysian business owner's full attention, because they describe exactly the habits I watch small teams skip when they first hand an agent something that matters.

1. Decide what "done" looks like before you press go

Azhar's first lesson is that agents are not lying when they announce a task complete while it is plainly unfinished; they are guessing at a finish line nobody actually gave them. His fix, which he says is worth doing on paper before touching the keyboard, is to make the finish line something the agent can test rather than something it must interpret. His own example is the difference between telling an agent to make a Rubik's Cube "look more organised" and telling it the cube counts as solved only when every face is one colour. The first instruction invites a plausible-looking guess; the second leaves no room for one.

Most small businesses skip this step entirely, because writing a testable finish line feels like more work than the task itself, and because we are used to giving instructions to people who already share our sense of what "good" means without being told. An agent shares nothing of the kind. It will produce something confident and wrong with exactly the same tone it would use for something correct, and the only defence is deciding, before you delegate, precisely how you or the agent will know the work has actually met the bar.

2. Not every task deserves your most expensive model

The second lesson is a change in the question worth asking. A year ago the natural question was which model is best for a given job; Azhar's team now asks where in a piece of work extra intelligence actually changes the outcome, because the answer is very often "not here". Their own agents run the cheaper, faster model for routine work by default, and only bring in a stronger one at the points that genuinely need judgment, such as framing what a piece of research should even mean before the grunt work of gathering it begins.

The article cites a telling benchmark: pushing one current model from low to maximum reasoning effort improved its score on a standard intelligence benchmark from 49 to 59, a real gain, but the last increment of that push, from high effort to maximum, doubled the tokens (and therefore the cost and wait) for a single additional point. The team's own rule of thumb, borrowed from an Anthropic briefing they attended, is to prefer a stronger model working lightly over a weaker one straining at full effort. For a small business, the practical version is simpler still: know which of your recurring AI tasks are genuinely difficult and which are merely repetitive, and stop paying premium-model prices for the second kind.

3. Judge the work an agent hands back, not the tokens it used

The third lesson is the one I'd put in front of any owner nervous about an agent subscription's monthly bill. A token count tells you what went into the machine, not what came out the other end worth having, and Azhar's team now runs what they call a light weekly audit of their agents' substantial work: what got attempted, what got used, what it cost, how long it took to brief and check, and what a person would have charged to do the same job. His own first audit found one week's agent work, sixty-two tasks costing around eight hundred US dollars, would have cost roughly nineteen thousand dollars and forty-eight hours of his own time if commissioned from people instead. He is careful to call this an estimate rather than an accounting-grade figure, which is the right amount of caution, but even a rough version of that audit is more informative than watching a subscription fee and wondering if it is worth it.

Why this matters more for a small Malaysian business than a large one

A team the size of Exponential View's can afford a week of sloppy agent management as a rounding error. A five- or ten-person Malaysian SME cannot: an agent turned loose without a testable finish line, running on the most expensive tier for work that didn't need it, with nobody checking afterward whether the results were worth what they cost, will burn through a modest AI budget and a great deal of goodwill inside the business before anyone stops to ask why. The three habits above cost nothing to adopt and very little time to practise. What they buy is the difference between an agent that quietly earns its subscription and one that quietly becomes the reason someone on your team decides AI "doesn't really work for us".

The Business Owner's Questions

How do you use and manage AI agents? I hear people saying they run hundreds of them, sounds too wonderful to be true. Are they free? Business owners may think these are embarrassingly simple questions, but they need to understand what agents are, how they work and the return on investment (ROI) they can get from them. They can also turn out to be financial minefields.

What This Means for You

None of this requires waiting for a more capable model or a bigger budget; all three lessons are about discipline rather than technology, and all three are cheaper to practise than to relearn after an agent has already cost you a client's trust or a week of rework. Write the finish line first, spend your most expensive model only where it changes the outcome, and look honestly, once a week, at what the work actually cost against what it was worth. Azhar's team learned this running agents at a scale most SMEs will never reach; the lessons are, if anything, more urgent at the scale most of us actually operate.