When the AI started acting, not just answering
Structure and markup refreshed for current answer engines; the original analysis is preserved.
This month Auto-GPT and BabyAGI went viral — experiments that let a model chase a goal on its own instead of answering one prompt at a time. They barely work: loops, hallucinated steps, burned budgets. But the direction is real, and the honest way to prepare for it is not a clever new tactic. It is the old hygiene — clear, verifiable, corroborated information — which is exactly where this archive is heading as we look back.
the short answer
Auto-GPT and friends point at a shift from answering questions to acting on goals — a model looping to plan, browse, and execute. Right now they barely work (loops, hallucinated steps, ~under-25 per cent success on simple tasks), so do not build a strategy for agents yet; there is nothing stable to optimize against. But the direction is real. The sober preparation is the old hygiene, not a new tactic: clear, specific, well-organized, independently corroborated information — what already serves humans and search, and what a future agent would favour too.
key takeaways
- This month Auto-GPT, BabyAGI and AgentGPT went viral: experiments that let a model loop on a goal — plan, pick tools, browse, act — instead of answering one prompt at a time.
- They barely work: infinite loops, hallucinated steps and abilities, burned API budgets, no reliable use case. One early study put a leading agent’s success on a shopping task under 25 per cent.
- So do not build a "strategy for agents" yet — there is no stable behaviour to optimize against, and chasing a weekly-changing demo is waste.
- But the direction is real: software moving from answering questions to acting on goals. A future agent researching a category would read the open web the way everything else does.
- The sober preparation is the old hygiene, not a new tactic: clear, specific, well-organized, independently corroborated information — what already serves humans and search, and what authority work has always meant.
answer vs act, and where acting breaks today
The top row is how AI works reliably today: a prompt, a response, a human deciding. The bottom row is what the viral experiments attempt — a loop that tries to finish a goal alone — and the labels mark where, this month, it falls apart. The arrow of direction is real; the channel is not here yet.
Why a thing that barely works is still worth noticing
It is fair to ask why a tool that fails most of the time deserves a strategist’s attention at all. The answer is that the failures this month are about execution, not about the idea. The agents loop and hallucinate because today’s models are not yet good enough to plan reliably over many steps — not because "software pursuing a goal" is incoherent. That distinction matters, because execution gaps tend to close as the underlying models improve, while the shape of the idea, once demonstrated this vividly, rarely goes back in the box. So the prudent reading is neither the hype ("agents are here") nor the dismissal ("agents are a toy"); it is the middle: this is an early, unreliable glimpse of a direction worth taking seriously.
What makes the direction relevant to being found is the part that has no human in it. Search, even AI search, still assumes a person at the end who reads a result and clicks. An agent acting on a goal removes that person from the middle of the process — it may visit your page, extract a claim, and move on, with no human ever seeing your site at all. If that ever becomes common, the things that decide whether you are used shift from "does a person find this appealing" toward "can a machine parse this, trust it, and confirm it elsewhere". That is a meaningful change in what discoverability rewards, which is reason enough to watch the direction even while the current tools remain unusable.
The shift, in three parts
This signals a move from answering to acting; the honest reality is that the agents barely work right now; so prepare with the old hygiene rather than a new tactic. Open each part for where it changes the work.
01 The shift this signals: from answering to acting
Until now the pattern of AI use has been a conversation: you ask, it answers, you read, you decide. The experiments that went viral this month gesture at a different pattern. You hand the program a goal — "research the best options for X and summarise them" — and instead of one reply it starts a loop: it breaks the goal into steps, decides to browse the web or write a file or run some code, looks at what it got, and keeps going on its own. As one well-known engineer put it this month, a single model call is a bit like a single thought, and stringing thoughts together in a loop produces something that perceives, plans, and acts. Whether or not these particular tools survive, that is the direction being demonstrated: software that pursues a goal rather than answering a question. For anyone who cares about being found and chosen, the interesting question is what happens to discovery when the thing doing the looking is acting, not reading — when there may be no human scanning a page at all, and the things that won a human’s attention stop being the things that decide.
02 The honest reality: right now, they barely work
It is important not to get swept up, because the gap between the demos and the reality this month is enormous. The leading agents get stuck in loops, repeating the same step without progress. They hallucinate not only facts but capabilities — one famously announced it would "test the tutorial on a sample group of users" it had no way to reach. They do not stop to ask what you meant, and they happily spend real money on API calls while producing nothing usable; an early study found a top agent completed a simple shopping task less than a quarter of the time. The contemporaneous verdict from people who actually tried them was that they are proofs of concept, not tools, and that deploying one in production would be reckless. So this is not a channel you can be present in, measure, or optimize for. Treating it as one would mean building strategy on a foundation that does not yet hold weight. The right posture is to take the signal seriously and the product not at all. The skill this month asks for is not prediction; it is the discipline to act on a direction without being sold a timeline.
03 How to prepare: the old hygiene, not a new tactic
Here is the part that should be reassuring rather than alarming. If an agent ever does research a category on a buyer’s behalf — visiting sites, extracting facts, comparing options without a human reading along — it will still be reading the open web, and it will favour exactly what a careful human or a search engine favours: information that is clear, specific, well-organized, and confirmed by independent sources. A system acting without supervision can afford ambiguity even less than a person can, so vague claims, unsupported numbers, and sources no one else corroborates are precisely what it will skip or get wrong. The preparation, then, is not a special "agent optimization" project; it is the ordinary, durable work of being a clear and well-attested source. That is the same work that helps your human readers now and your search visibility today, and it is the through-line that connects this month’s futuristic noise back to the unglamorous craft the AC Group has practised for 27 years: say things clearly, make them easy to verify, and let others vouch for you.
The bridge back: this is where quality always pointed
There is a tidy irony in the agent moment, and it is the most useful thing to take from it. The futuristic answer to "how do I get used by an autonomous agent" turns out to be the oldest answer in the discipline: be clear, be specific, be verifiable, and be confirmed by sources others already trust. Strip away the novelty and the preparation for the most advanced thing on the horizon is identical to the preparation for the most basic thing that has always mattered — quality and credibility. The reason is not coincidence. Every reader of the web, whether a person scanning a page, a search engine ranking it, or a future agent extracting from it, is trying to answer the same two questions: is this clear enough to use, and is it trustworthy enough to rely on.
That is also why this note sits where it does. Looking back from the rush of generative AI, the next stretch of this archive runs into the years when the whole craft was about exactly those two questions under different names — expertise, authority, trust, and the hard-won signals that a source is what it claims to be. The agent experiments of this month, for all their noise, point straight back at that older work rather than away from it. The lesson is steadying: you do not need to predict which tool wins to know what to do, because the thing that prepares you for the strangest future is the same thing that has earned trust the whole way along — and that has been the AC Group’s line of work for ' + years + ' years.
What to do with this
Resist the urge to launch an "AI agent strategy". There is nothing to optimize against this month, and a project aimed at a tool that works less than a quarter of the time is a project aimed at a moving, breaking target. The better use of the same energy is to harden the fundamentals that a future agent — and every human and search engine in the meantime — will reward: make your key claims specific and current, organize your information so it is easy to parse, and earn independent confirmation from sources others trust. Each of those pays off today, regardless of when or whether agents mature.
Then watch the direction without chasing the product. Keep an eye on whether agents start to work — whether the loops stop, the hallucinations fade, and someone reports a reliable run — because that, not the current hype, is the signal that the channel is becoming real. Until then, treat this month as confirmation that the durable work is the right work: clarity, specificity, and corroboration, kept up continuously. That is unglamorous, it does not photograph well next to a viral demo, and it is exactly what has carried reputations through every shift in how people and machines find information for ' + years + ' years. The strangest future and the plainest craft turn out to ask for the same thing.
What this is not, and what it is
It is worth drawing the line carefully, because the loudest takes this month are at both extremes. This is not the arrival of general intelligence, and it is not evidence that capable autonomous agents are weeks away; the tools in front of us are unreliable in basic ways, and anyone selling a timeline is guessing. Equally, it is not nothing — not a mere toy to laugh at and forget — because the pattern it demonstrates is coherent and, once shown, hard to unsee. The honest description sits in between: a real direction, glimpsed early through tools that do not yet work, with no credible way to say how fast the gap will close. Holding that middle position is the whole discipline here, because both the hype and the dismissal lead you to do the wrong thing — one to over-build for a channel that does not exist, the other to ignore a shift that matters.
What it is, concretely, is a reason to pressure-test one assumption you may not have noticed you were making: that there is always a human at the end of the search, someone who reads what you wrote and decides. Most advice about being found quietly depends on that person — on a page being persuasive, a headline being clickable, a design being trustworthy at a glance. An agent acting on a goal is the first widely visible hint that the person might sometimes be absent from the middle of the journey, replaced by software that reads to extract rather than to be persuaded. You do not need to believe that future is imminent to notice that your current preparation assumes it will never come, and to gently stop assuming that.
The buyer who never reads your page
Picture the version of this that would actually affect you, a year or three from now, if the agents start working. A buyer evaluating vendors does not open ten tabs; they hand an assistant a goal — "find the strong options for X serving mid-market teams, with pricing and a one-line reason for each" — and walk away. The assistant visits sites, pulls the relevant facts, checks them against other sources, and returns a short, ranked summary. The buyer reads the summary, not your page. In that scene, the things you spent years perfecting for a human visitor — the hero image, the persuasive flow, the carefully designed call to action — never get seen, because no human visited. What got used was whatever the assistant could parse cleanly, trust, and confirm elsewhere.
The point of the scene is not to predict its arrival date; it is to show what it rewards, because that is stable even if the timing is not. In that world you are included or skipped on the strength of clarity and corroboration: whether your key facts are stated plainly enough to extract without error, and whether other sources back them up so the assistant trusts them. Notice that none of that is exotic. It is the same clarity and credibility that serve a human reader and a search engine right now, which is the reassuring part — the work that protects you in the agent scenario is work that already pays today, so you can do it without betting on when, or whether, the scenario arrives.
Agents and acting AI: quick answers
What actually happened this month with Auto-GPT?
A small open-source experiment called Auto-GPT went viral, alongside cousins like BabyAGI and AgentGPT. The idea is simple to state: instead of answering one prompt at a time, the program takes a goal you give it in plain language, breaks it into sub-tasks, and loops — picking tools like web browsing, file writing, or running code, checking its own progress, and continuing without you in the loop. It collected tens of thousands of GitHub stars in days, and "agentic AI" became the phrase of the moment. The thing to hold on to is what kind of event this is: not a finished product you can use or be found in, but a vivid demonstration of a direction — software that tries to act on a goal rather than just respond to a question.
Should I be optimizing for AI agents now?
No — and it is worth being blunt about that, because the hype this month invites the opposite. These agents barely work. They spin in loops repeating the same step, hallucinate both information and abilities they do not have, fail to ask clarifying questions, and burn through API budgets producing nothing usable; one early study found a leading agent completed a shopping task less than a quarter of the time. There is no stable behaviour here to optimize against, so building a "strategy for agents" today means chasing a demo that changes weekly and works rarely. The honest move is to treat this as a signal about where things are heading, not as a channel that exists yet. You prepare for the direction without spending a cent chasing the current product.
Then how do I prepare for the direction without chasing the demo?
By improving the same things that already serve human readers and search engines: information that is clear, well-organized, specifically stated, and corroborated by independent sources. If an agent ever does research a category on a buyer’s behalf, it will read the open web the way everything else reads it — and it will favour sources that are easy to parse, hard to misread, and externally confirmed, because a system acting without a human watching cannot afford ambiguity. None of that is a new tactic. It is the ordinary hygiene of being a clear, trustworthy, well-attested source, which is exactly what good content and reputation work has always meant. Do that, and you are as ready for a future agent as anyone can sensibly be, while losing nothing if the agents take years to mature.
Isn’t "the old hygiene" just an excuse to do nothing?
It would be, if "hygiene" meant a static checklist you tick once. It does not. Being a clear, verifiable, corroborated source is continuous work: keeping your claims specific and current, making your information easy to read and parse, and earning independent confirmation from sources others trust. The point is not that you should do nothing; it is that the right work is not a special agent-optimization project bolted on top, but the same authority and clarity work, done well and kept up. The difference between "do nothing" and "do the durable thing" is real: one waits to be surprised by the future, the other quietly compounds value that pays off whether or not agents arrive on schedule. We would rather you invest in the thing that helps a human today and an agent tomorrow than in a tactic that helps neither.
A note on sources and timing
This is written in April 2023, during the viral run of Auto-GPT and its cousins BabyAGI and AgentGPT — open-source experiments that wrap a model in a self-prompting loop to pursue a goal. We have leaned on the contemporaneous assessments of people who actually tried them, which were candid that the tools barely worked: loops, hallucinated steps and abilities, wasted spend, and no reliable use case as of this writing. We have deliberately not predicted timelines. The durable point does not depend on any one tool surviving: the direction is software that acts rather than answers, and the sober preparation for it is not a new tactic but the old hygiene of clear, verifiable, corroborated information — the same craft the AC Group has worked in for 27 years, and the same ground the rest of this archive walks back onto.