# Foomers vs doomers: skilled incompetence and the AI tug of war > Source: > Published: 2026-08-03 00:00:00+00:00 # Foomers vs doomers: skilled incompetence and the AI tug of war By [Iain](/blog/iain/), : [AI](/blog/ai/), [Process](/blog/process/), [Hot takes](/blog/hot-takes/) foom-er, noun, an ironic malapropism playing on doomerism (AGI fatalists) and FOMO (hustle bros), used to describe techno-optimists. Reflects the onomatopoeic “slowly, then all at once” nature of exponential growth: “AI go foom.” A speaker at one of the many AI conferences that have sprung up like garments on your “clothes chair” (not clean, but not dirty enough to wash) describes the miracles of agent-assisted coding. Year-long rewrites finished in a fortnight, backlogs cleared, and hard problems dissolved by brilliant prompting and clever code harnesses. The audience of true believers laps it up, with some already drafting an email to their CTO with suggestions. What the speaker does not mention, because she may genuinely not know, is that several teams at her own company are still grinding through the cleanup from those very same miracles months later. That gap between the conference case study and the messier reality is now the defining tension in most organisations using agent-assisted coding at scale. In some, it has hardened into something like a religious schism, with boosters and sceptics firmly entrenched in opposing camps. Although the split looks cultural, it is not. It is a systems failure: a feedback loop that has broken in specific, diagnosable places, and a broken loop is a design problem with possible design solutions. ## Both sides are racing, both sides are right The enthusiasts are racing against the clock. Teams that have adopted AI-assisted coding are seeing real, discontinuous jumps in output, and those jumps do not look like a typical technology cycle, where the cautious can wait for the dust to settle. Sit it out, they reason, and you risk obsolescence. The sceptics are racing entropy, the steady decay that follows when code ships faster than anyone can read it, into systems where no single person holds the full picture, drawing down on a trust account that took years to establish. Neither side is imagining the threat. Across [more than a hundred thousand developers on GitHub](https://www.nber.org/papers/w35275) (NBER Working Paper No. 35275, Demirer, Musolff, and Yang, 2026), coding agents produced an 180% cumulative increase in coding activity and a 30% increase in actual releases. The engine is running much hotter, and a larger share of what it produces reaches the end user, but the ratio between the two has widened sharply. Eight times the raw output, a third more shipped product. The enthusiast is right that the acceleration is extraordinary. The sceptic is also right that the consequences of the acceleration are being absorbed somewhere, and “somewhere” is usually the on-call rotation. ## The soggy sandwich Arvind Narayanan and Sayash Kapoor, in [ AI as Normal Technology](https://www.normaltech.ai/p/why-ai-hasnt-replaced-software-engineers), argue that AI spreads through the economy gradually because some parts of work resist automation and remain stubbornly human. Fred Brooks identified the deepest reason for this forty years ago in [(1986): the hard part of software was never the typing. It was figuring out exactly what the thing should do.](https://www.cs.unc.edu/techreports/86-020.pdf) *No Silver Bullet*Their framing of the work is as a sandwich. The middle layer is execution, where the code is written. The two outer layers decide what to build and who owns what ships. And owning means testing, review, and the long slog of keeping the thing alive in production. Agents have compressed the middle layer almost to nothing, but have had significantly less impact on the outer layers. This is not a new pattern. W. Edwards Deming spent his career arguing that optimising parts of a system independently will sub-optimise the whole, and that the damage is invisible until you measure the system end to end. An engineer who quintuples her commit rate is optimising one subsystem. The release pipeline, the on-call rotation, and the review queue are other subsystems, each with its own capacity limit, and none of them will improve because of her. Perhaps, quite the opposite. Understanding the code is still an unresolved bottleneck. There is a convincing argument that understanding is only truly gained during the writing loop. That the understanding you gain from just reading or code review is marginal compared to the understanding you gain while writing. That this time contributed to updating and deepening understanding of the system under development. That there’s no replacement for that understanding in a world where LLMs are doing the writing. The extent to which this truly matters, both now and as model capability repeatedly jumps forwards, is contested ground. ## A principal-agent problem wearing a hoodie The structural reason for the cost shift will be familiar to economists, but less so to Scrum Masters. It is the principal-agent problem, which arises when the person making a decision does not bear the consequences. The developer who ships the AI-accelerated feature gets the high-five. The engineer who discovers at 2 am that the feature has been silently corrupting a cache for a fortnight does not get one. In a large organisation, the original developer may never learn that the two events are even connected. Most software engineers, if you catch them after a few drinks, will offer some version of the same confession. They do not want to be the AI hater. They studied this stuff, find it fascinating, and feel backed into a corner where they have to play the killjoy because they are the last person in the room asking hard questions about tools everyone else has already decided to love. Chris Argyris would have recognised the dynamic immediately. He called it “skilled incompetence,” the phenomenon in which smart, well-meaning people learn to suppress uncomfortable data to avoid appearing obstructive, until the organisation loses the ability to hear bad news altogether. The sceptic tends to self-censor. The enthusiast never learns the knotty consequences. Both are practising single-loop learning, adjusting their behaviour within a fixed frame, when what they need is [double-loop learning](https://hbr.org/1977/09/double-loop-learning-in-organizations), the kind that steps back and asks whether the frame itself is correct. Argyris was describing self-censorship, but I’ve also heard from several fellow consultants that questioning the reasoning for a project is for or expressing visible doubt about a client’s AI strategy has become close to a fireable offence in certain organisations. Some engineers have even adapted by “AI-washing”. They do the work the way they always did, then report that Claude did it, because the manager is unhappy when the tool does not appear in the process regardless. Although anecdotal, the behaviour is what Argyris predicted. When the frame cannot be questioned, people stop supplying the data that would question it and get inventive about the appearance of compliance. ## The four places the circuit breaks Strip away the human emotions, and there are four specific structural failure points: ### The true cost is unmeasured Improvements usually arrive with a number attached: the migration shipped in 100 fewer hours than before, the backlog slashed in days. The cost shows up as a vibe, a grumble in a retro, a complaint over DM. One is data, and the other is anecdotal. ### The path home is severed The cost usually lands on a team that did not create it, so there is often no way to trace it back to the person who made the original decision. Cause and effect have been uncoupled. This scales with organisational size; the bigger the company, the less likely a reliable feedback loop exists. ### The clocks do not match Wins are immediate, while costs may lag by weeks or months. By the time the cleanup bites, the celebration is old news, and nobody makes the connection. ### The forums are asymmetric Wins get the high-status channels: the all-hands mentions, the conference talk, the launch post. Costs tend to get the low-status ones: the SRE meeting, the on-call handover, the grumbling in standups. ## Rebuild the circuit These are accidents of structure. So if the loop is broken in four places, the obvious fix is to repair it in four places. None of this asks anyone to like AI any more or less. It simply asks them to reconnect cause and effect so the organisation can learn, which is what a feedback loop is for. Charity Majors, in the second edition of [ Observability Engineering](https://www.oreilly.com/library/view/observability-engineering-2nd/9781098179915/) (with Liz Fong-Jones and George Miranda), puts it exactly: timely, precise, and relevant feedback loops produce self-awareness in people and self-governance in teams, without anyone standing over them. That is the standard to aim for. **Surface the cost.** The sceptic’s complaint stays invisible only as long as it remains a feeling. Count the cleanup hours, the defect backlog, the incident load, and tag those that trace back to AI-accelerated work. The model here is [Fin’s own write-up](https://ideas.fin.ai/p/2x-nine-months-later), in which Darragh Curran reported that his R&D org tripled its output in nine months and that code quality declined for eighteen frightening months before eventually recovering. He published the dip alongside the triumph because he was measuring both. One cheap instrument is worth using. Poll the team anonymously and ask each person to rate the project’s chance of success out of ten. My experience is that the typical shape is bimodal, half the room at three and half at eight, on work that is already late. An average would bury that. The distance between the two clusters tells you how much is going unsaid. Decide early and clearly what counts as AI-accelerated work, because the label is already being gamed. I see many current AI projects that are just ordinary projects with an AI component bolted on to clear an internal hurdle or secure budget allocation. **Route the signal home.** Tag AI-accelerated changes so the downstream cost can be traced back to the decision that caused it, not as blame but as open, visible information. This is Argyris’s suggestion, and it is the most culturally difficult, because it requires the organisation to surface data that makes people uncomfortable. Most humans and therefore organisations (collections of humans) are very good at learning what they already want to believe and very bad at learning anything else. There is a layer above the team where this gets harder. Very senior executives can be the biggest Foomers of all, and the most divorced from operational reality. They may already have publicly claimed implausible productivity gains, despite a chain of people beneath them holding private doubts that can be career-limiting to voice loudly. If that is the executive weather above you, tracking real AI-accelerated changes is a local act of sanity rather than a fix. but it is still worth doing anyway. A team that knows what its own work costs has something solid to stand on when the numbers coming down from above stop resembling anything rational. **Match the clock.** A celebration on launch day closes the books long before the real balance appears. Add a check thirty or sixty days later, tied to the same change, to ask what it actually cost or saved overall. The launch is the first chapter of the story, not the last. **Put the cost in the same room as the win.** Make the enthusiasts acknowledge, with every win, how much anyone downstream may be mopping up, and make the sceptic raise the cost in the forum where the win was first announced rather than in a private channel. The asymmetry dies when reporting the true overall cost carries the same social standing as reporting just the win. ## Turn the fight into a precondition Once both sides agree on the numbers, the argument changes. I have [previously written about the many forms of AI-assisted coding](https://betterthangood.xyz/blog/vibe-coding-spectrum-weekend-hacks-dark-factory/), ranging from pure vibe coding to proper agentic engineering. The sceptic’s real fear is the first being passed off as the second. Majors’ research describes a typical scenario in which an impasse emerges. An enthusiast wants to ship without code review because they heard another company does it without problems, while the sceptic flatly refuses to auto-accept changes. The productive version hinges on a single question: what would it take for someone to feel comfortable shipping to production without reading every line of code? Better evals. Better tests. Feature flags, blast-radius reduction, and a first attempt on something well off the critical path. The disagreement then becomes a backlog, and an engineering team always knows how to work a backlog. ## The amplifier with no neutral setting Nathen Harvey’s memorable line in the 2025 DORA report compares AI to an amplifier. It magnifies the strengths of a disciplined organisation and the dysfunctions of a struggling one, and there is no neutral setting on the dial. Fin tripled its output because it already ran tight feedback loops and measured everything. The tooling found that discipline and multiplied it. Point the same tooling at a team that ships on vibes, and it will reliably multiply them as well. A person also earns moral authority by owning the consequences. If you are the one left holding the bag, you should have a say over what goes in it. The engineer who supports the code should help set the terms under which it is created and shipped. But that standing has to be earned under the new terms, not the old ones. A critique mounted entirely from inside the way things used to work has no purchase on a conversation about how they work now. The sceptics who will actually shape how AI gets used are those who become credible about AI, who can say with a straight face that they see the opportunity and want to help everyone reach it safely, and who therefore gain the standing they need if and when they pull the safety cord. Like this? [Get email updates](https://betterthangood.xyz/newsletter/) or [grab the RSS feed](https://betterthangood.xyz/feed.xml) like it’s 2008. ## More from the blog - [Forward Deployed Engineers: AI’s hired guns](/blog/hired-guns/)The last mile of AI deployment is a real person, sitting in your office, learning your business, writing your plumbing, and making things work reliably. That will remain human for the foreseeable future.Google Trends shows a 5,000% increase in searches for “AI forward-deployed engineer” over the past five years. The curve stays flat from 2021 through mid-2024, then turns vertical. Indeed data reported by Business Insider shows postings for the role rising from 643 in April 2025 to 5,330 a year later, a 729% increase. LinkedIn’s own study found 42-fold growth between 2023 and 2025, making fo… - [The sorting machine: recruitment’s race to the bottom](/blog/the-sorting-machine/)The sorting machine was built to cope with a genuine problem. Where it went wrong was in mistaking efficiency for effectiveness, in assuming that the ability to process resumes fast meant the ability to evaluate candidates well. The fix is not a smarter machine but a more modest one, paired with humans who are given the time, the tools, and the structured information to do what they were always better at.If you’ve been involved in any part of the recruitment process, the last 18 months have been dispiriting, to say the least. Hiring became a closed loop. Employers adopted applicant tracking… - [Weighing smoke: why GEO dashboards are mostly useless](/blog/weighing-smoke/)The promise of a tool that gives you similar search visibility to the pre-agent world is understandably seductive. But before you eagerly hand over your cash for a platform, consider whether an afternoon’s work a month might be a more cost-effective alternative.I recently wrote about the GEO chimera, the cottage industry promising to optimise your brand into AI answers. My argument was that the levers being sold mostly did not exist. Another part of the industry is selling gauges rather than levers. Estimates of the spend on AI visibility tracking now run past $100 million a year, across hu… - [Another nice mess](/blog/another-nice-mess/)Somewhere in your business right now, someone is assembling a picture that no single app can provide. It may be the project manager pulling hours from Harvest and budget data from the finance tool to assess whether the engagement is still viable. Maybe it’s you on a Sunday, because what you need is not any one number from a system, but the pattern across three of them. The cloud gave small businesses access to the best software they had ever had, priced monthly and built for specific purposes. But twenty years of sensibly chosen apps have left the average small business with a patchwork data … - [The state and the machine](/blog/the-state-and-the-machine/)What little we saw of Fable and Mythos offers both cause for excitement and concern. It was widely and credibly seen as a model of a completely different caliber from those that had come before. Perhaps the risks in this instance were overstated or amplified for political ends. What is more profound is that the short time we had with the models offered a clear glimpse of a future in which a single company is making significant progress toward a superintelligence with the potential to rival or exceed the power of nation-states or even massive corporations. That juncture was never going to ar…