> ## Content Index > Fetch the complete content index at: https://bayesiantheory.com/llms.txt > Use this file to discover other available public pages before exploring further. # AI agents: the future or a fugazi? - URL: https://bayesiantheory.com/ai-agents-the-future-or-a-fugazi/ - Published: 2026-08-28T09:27:49.000Z - Updated: 2026-10-05T14:45:11.000Z - Description: If agents are so great, why can’t we get them to work? - Author: Gregg Bayes-Brown - Tags: AI Safety, artificial intelligence, AI risk, agents, Mark Zuckerberg, agentic, AI, psychology, facebook, Newsletter, #Migrated-1791211413550, #Import 2026-10-05 14:44 `Image credit: JZingerman` When it launched in August last year, OpenAI’s agentic coding harness Codex instantly sent me spiralling into the depths of AI vampirism in the best possible way. I hadn’t had this much fun on a computer since the launch of Grand Theft Auto 5. It’d take my best mate texting me saying “you do know you haven’t seen anyone in six months?” to snap me out of its trance. One pub trip later, and I booted Claude Code for the first time. Rinse and repeat. A year into the agentic era, and it’s clear I was far from the only one. But underneath the collective hype, a voice has been getting louder and louder: are these agents actually doing anything useful? Indeed, are they even the right solution to the problems we’re trying to get them to solve? [Subscribe now](#/portal/signup) ### The name’s Bond, json-bond Before we can answer that, a quick explainer. A good catch-all for agents is essentially AI that can carry out tasks on your behalf. Unfortunately, this is also where the agreement on what an agent actually is begins and ends. In my understanding, I lump them into two categories – robots and mechs. The former is a brain-with-scaffolding format designed to work autonomously aside from humans. The latter requires a human and machine working together: the agentic harness. Think Claude Cowork/Code, or Codex/GPT Work - incredibly powerful, but nothing without a pilot. I’ve experimented with a vast array of agents now. While mechs are what I use day in, day out, it’d be an overstatement to say they’ve made robots redundant. Robots are still the types of agents offered by countless folks online and deployed at scale by forward-deployed engineers (FDEs), and you can get great utility out of them. Some uses I’ve found for robots so far include: - Intel bots which scour the net for news, opportunities, & risks for organisations. - Grocery bots which plan meals, design shopping lists, and puts in the order. - Pre-meeting briefs, dynamic calendar management, and post-meeting transcript into documentation. While you can get fantastic utility out of these (Morgan Stanley reckons they get about 15 hours back a week for client-facing staff for just the post-meeting part of the last one), it’s rarely the smooth sailing the world’s GTMs and FDEs would have you believe. But don’t take my word for it: ask Mark Zuckerberg. ### Facecooked Meta and poor agentic use are well established bedfellows at this point. I, like many, have been unceremoniously exiled from Zucktopia by the company’s earlier dabbling with agents. The number include many business owners which have grown substantial followings on Instagram or Facebook. One day you’re there. The next, you’ve been Kaiser Soze’d. Like the rest of GenAI, Meta’s banhammer is a probabilistic system. Say you’re a CEO known at the local golf club as Billy Bigballs for all the people you culled in the last quarter. You’re looking at your customer service team as fresh meat and go all Agent of Death. Your agentic call centre staff get a 98% successful call completion rate. Because its probabilistic, 2% go haywire. Yet, that’s better than the 94% your carbon-based call centre staff could achieve, so trebles all round after the 18th hole. The same logic does not apply in workstreams where the 2% matters. Back at the start of the year, The Zuck began Project OT: Meta’s big bet on going AI native. Like Billy, Zuck’s definition of AI native meant blood in the streets - with a planned a cull of 60% of Meta staffers. He planned for a two-wave assault, with the first wave of 10% getting the axe in May. Zuck had to cancel part B. Say what you will about the company (and if you don’t, I will), but Meta doesn’t typically hire idiots. As such, the Metaverse noticed what was going on and revolted. But that’s not the whole picture. The more crucial factor is that the agents themselves. They were unreliable, brittle, and easily went rogue. But it’s not when agents fail that’s the real problem. It’s when they work. ### Feel the burn For the entirety of my time as an AI vampire, I was teetering on burnout. The fact I didn’t is more down to my neurodiverse journey– and the stimulants it provides – than anything anyone sane would deem as healthy work/life balance. AI’s speed makes my amphetamines look like sedatives. As I wrote this, Bill Gates – the literal poster-boy alongside Jobs of disruption in the dotcom era – said we’re going too fast. For me, it’d be like AC/DC saying I rock too hard – an inconceivable intervention, but one I would ignore at my peril. Nearly all the AI folks I speak to mirror my issues with burnout, and a chief culprit is the speed. Meta was no different. Agents can massively increase productivity – and with that comes increased output. For every one token burned by a human, agents are burning five. But as discussed, this isn’t a deterministic system here – it’s probabilistic. That means someone’s got to check that for that 2%. This is tough enough to do when you’ve got the necessary AI skills and domain knowledge, but this is a rarity. For previous technologies, this wasn’t a problem. You don’t need to know how a computer works to use it effectively. But, as Gates pointed out, AI isn’t like previous technologies. Missing that fact is precisely the trap of “AI is a tool” I spend my days rallying against. The barrier here is our ability to process outputs. Even a great output is rendered useless when the user does not understand them. It’s not just AI which requires context – we do too. The probabilistic nature adds cognitive burn. We’re not just looking to understand the output, but catch the 2% too. When multiplied through increased agentic output, multi-tasking, and context switching, even the best of us will inevitably default to automation bias and start letting our guard down. Consequently, AI is the first technology which requires theory of mind to use safely and effectively. The user is both where all the risk sits, and where all the opportunity lies. To mitigate the former and maximise the latter, psychology and sociology should be the bedrock of any enablement programme. Unfortunately, this is not what most organisations have done. To make up the skills shortfall, many businesses instead turn to FDEs and third-party vendors – essentially offshoring control of core business processes. While this might get you an agent in the short term, it fails to build up in-house knowledge of the agents these engineers install. When the end user did not build the agent, they likely don’t understand it. This is the same comprehension problem of outputs now at the agentic layer. Further, they almost certainly cannot fix it when it breaks, or tweak it to refine it. The end result not only amplifies risk, but turns work into a miserable experience, where a human’s contribution is reduced to factory-line quality assurance. AI works very much as a mirror. It is hyper-individualistic in nature, taking the user’s intent, mindset and skill and turning that into reality. As such, agents work best when they are an extension of the user – built by the user bespoke exactly to the problem the user understands better than anyone else. By contrast, organisations are collectivist efforts. Any organisation which hasn’t considered culture, a shared narrative, and ways of working collaboratively inevitably discovers the consequences of not being joined up. When agents or indeed AI in general are allowed to run wild without establishing ground rules, hyper-individualism turns silos into towering monoliths overnight. For organisations, this becomes single points of failure dotted across their entire landscape. Supercharged personalities in conflict, incompatibility of outputs, and unclear processes only lead us down one path: organisational burnout. ### Keep it simple While I don’t have all the answers to these problems, a few things have become clear. First, never reach for a robot when a mech is available. Cowork/GPT work etc give individual users agency and keep them in the driving seat, increasing the odds they will remain cognitively engaged enough to make AI outputs genuinely useful. Next, when working together at an organisational level, it is frankly an unacceptable failure of accountability on the part of the employer to not provide what their employees need. This means continuous development of AI fluency and literacy, ringfence time in work to develop AI skills and understanding, and the provision of the frameworks for co-creation and collaboration in an environment which fosters safe and open innovation. Finally, don’t use an agent when you could just use a workflow. Productivity is not the measuring stick you need – impactful outcomes which moved the dial is. Quality over quantity, system 2 over system 1 thinking. When we adopted computers, organisations only moved as fast as the slowest typer in the company. With AI, we’re all the guy who’s been typing with their index fingers for the past three decades. For us to make a success of agents, our effort needs to placed on empowering that person, not attempting to run at the speed of Skynet. --- Thanks for reading. AI’s ever evolving, as is our common understanding of it. If I’ve got anything wrong here, please do call it out - that’s how we all improve. Would also be great to hear from folks who have similarly deployed agents - what’s worked for you, and what failed spectacularly? [Subscribe now](#/portal/signup)