As Purplelab's resident theoretician, I've been enjoying the saaspocalyspe discussion and questions of its degree, as a chance to dig into the literature. In matters of AI I've borrowed heavily from Henry Farrell, a Johns Hopkins theorist. His writing two years ago on the effect of LLMs was prescient, particularly of the applications for knowledge graphs and summarization, "LLMs’ impact is going to be most profound in the routine applications of culture – not traditional literature, but what J.G. Ballard used to call 'invisible literature' – the routine writing that holds organizations together and communicates their objectives to their employees and the world. The output of LLMs will likely replace much of this literature, and do a better job than it can do in helping coordinate activities." We have seen this play out internally and across the industry, where connecting agents atop knowledge bases have become a core part of finding and routing answers for both commercial and operational matters has substantially lowered the time to action, improving individuals efficiency and to a limited degree removing humans from purely information-transfer workflows. Like all the smarties say, it ain't coming for your job it's coming for each individual thing you do one by one, til you must go back to elementary school fractions and question what share of your job those functions are to tell if you'll still have a job by the end of this.
Of course that modality of AI use is beholden to the human in the loop framework which has become market standard for retaining trust in the outcomes of the machine process, and has served as a source of comfort for those thinking of long term job prospects. That framework seemed the dominant paradigm of AI adoption for most of last year, though the MCP and agent to agent protocols major players released clearly indicated the higher valuation application unlocks would not be merely "augmenting person" but automating workflows wholesale. The release of Anthropic's Opus 4.6 model more or less shattered that paradigm by vastly expanding the competence of models to operate independent of user. Not to heap undue praise on one lab- the pace of this industry means there are two near-parity competitors, and doubtless another will tip the performance a little further forward one of these days. But Opus takes the cake for getting a critical mass of builders, developers and founders, to embrace the use of the tools to the point where a new paradigm will form.
Openclaw/Clawdbot/Moltbot is an early indicator of this, though a privacy and security rats nest. If you didn't get the memo on why the nerds in your life were buying macbook minis, Clawdbot is why. Essentially it can connect via MCP to your inbox, your calendar, your bank account, you name it, and execute tasks you tell it to do. It may also delete all your emails if it misunderstands when you tell it to clean your inbox. The embodiment of humanity's lasting desire for a secretary, with some psychological hangover of a mother to it. Essentially though it is a tool for a person, which makes it part and parcel of the wave I think is cresting.
Here of course I'm reciting from Thomas Kuhn's Structure of scientific revolutions, the foundational text of understanding scientific epistemology, the frameworks through which we understand and interpret things. The new paradigm that's forming for business has a very clear structure I see from those at work building it. It does not resemble human in the loop frameworks in the slightest. AI in the coming year is being implemented operationally to solve orchestration problems across systems, and for R&D to build things, quite well enough on its own.
I'll dial in on the R&D one because it's more interesting and you'll need to be a client to get the goods on the former. So the structure that's coming to dominate is not "an agent, with sufficiently grounded context and background, to build a product or analyze a problem set". Perhaps it's inevitable for an industry that owes such a debt to parallel processing and unstructured learning, but it looks more like a team of agents, like a charioteer and a set of horses. The term "harness" has recently come into vogue to describe the set of controls, context, and tools an agent is surrounded by, like a horse in a harness. The trick to maximizing the performance of an output assigned to an agentic process is to add a meta-agent atop this team of horses, armed with the context of the end objectives, results of each sub agents operations, and a target output to assess completeness of the task against. An individual agent might have a range of fail-conditions – it could exhaust its compute budget, or the performance of it's output could be undesirable. So have the meta-agent assign the task to five sub agents, perhaps with different foundation models running, with iterations of sample data or suggested methods to complete the task. Then have the meta-agent, your charioteer or governor, whatever you'd like to call it, assess the comparative performance of the set and select or combine methods of the best answers. This shape, combined with reducing computational costs driven by data center development and chip improvement, suggests itself as the fastest path to successful R&D applications of AI in the near to mid-term.
In the long term, we're left to consider the limits of what this tech can do, which is why I was so happy Dr. Farrell'd recent book review / essay covering The Irrational Decision by Ben Recht (now on it's way to my door). Every paradigm has limits and every model has efficiency thresholds, and the present moment and opportunity is well illustrated by the following excerpt:
"Very large chunks of Silicon Valley’s current business model involve taking complex situations that don’t look like optimization or prediction problems, simplifying and redescribing them and then finding solutions. Just like statistics, there is a “sweet spot” for machine learning. It is not useful for situations where you have a genuinely clean mathematical abstraction, which you can turn into running code. Nor is it useful for situations that are too messy or complicated to be predictable (it is, after all, an application of statistical technique). You want to use it in the intermediary situations where there isn’t an obvious neat solution, but where the clunky and computationally expensive techniques of machine learning can discover a useful approximation, even if you may not understand quite what it is based on or how it works."
Some things don't require AI – a calculator for instance, will use much less energy and effort to arrive at the right answer to 2+2. Unit testing, the process of validating newly written code will perform as expected, is a great use case for ai. We are in the present moment figuring out which of those elements of business, previously locked behind unstructured language and across systems, are reducible to optimization and prediction problems, and testing whether an autonomous solution can arrive at better answers faster than a human alone.
Which leads me to a final thought- who will win out of this. I was struck by a recent piece in business insider on meta's change of org structure to flush out middle management and make leaders more "player coaches". It seemed to chime with the currently en-vogue valley term "high agency", used to refer to people who seem to be able to make things happen, doers of stuff. I don't go in for all that, but as I'm looking at the KPIs I'm setting for myself and my team, I did find this rang true:
"'Companies also need to set new performance goals and new ways to measure and reward success (in integrating AI into their practice).'It's all about implementation,' Pozner said. Otherwise, 'people are just going to do the same things that they've always done.'"
I suspect a flurry of dumb KPIs around using chatbots will be issued for teams across industry. The right answer to me seems to be floating point KPIs- learning and testing as much as using the current version. Let me know if you've written some good ones, I'm overdue to turn mine in.
Originally published on LinkedIn.