All Newsletters7 October 2026

The Agentic Evolution

Intro

The introduction of agents has been ongoing for around half a year. The technology of AI is now proven. It works, and we will prove this to you within this Newsletter. The positive investment returns from Ai for the Model Labs and the Hyperscalers are are now apparent to all but the most myopic (see later section). The actual issue for the Hyperscalers is that the returns are so attractive they have become rationally obsessed with “build baby, build” to take share from their competitors via escalating the cadence of their GW roll out.

This has led to a crowding out of capital markets which is playing havoc with an already over indebted, fiscally undisciplined Treasury market. The model Labs are tweaking their models after realising that consumers don’t want to pay for Ai subscriptions, but enterprises do want to pay, and the Labs have learnt direct token charging is more attractive than subscriptions. The Labs are also learning that they may have to sell a good proportion of their enterprise offering through third parties as Main St feels more comfortable consuming their tokens via their Hyperscaler platform.

The Key Risks for the Era of AI as they stand now are:
Regulation (including P(extinction) concerns): many US States are already making Data Centre construction almost impossible.
Power Constraints: on grid does not have enough capacity, and off grid is supply constrained & facing environmental push back.
Adoption: that the adoption of agents and token growth stalls because many Main Street businesses find adoption too difficult or time consuming.

In this Newsletter, we are going to take a look at the Adoption risk from a micro angle.

“What I cannot create, I do not understand.” Richard Feynman

As a digression, this quote was found written on his blackboard at Caltech shortly after his death.

The smart modern investor no longer asks of an enterprise: “how many employees do you have?” They ask: “What is your employment spend and what is your agentic token spend per annum?” Many lower tier financial analysts fail to recognise that the most profitable financial institutions in the world have the highest profit to employee ratio precisely because their business is built on models, quants and algorithms, not humans. Hence the question that high quality Investment Fund Analysts should be asking is: “I know you have x humans in your Investment Team but can you give me your latest annualised Token spend by Agents utilised specifically in the Investment Process.”

For MLCM, our spend specifically on the Investment Process (so monitoring financial data, analysing stocks, building Quant portfolios, trading strategies etc.) is now running at over £100,000 per annum. We would guess that in 3 years time, MLCM will be spending at least £500,000.

We have been actively building and testing agents, to undertake a number of enterprise functions from Invoice accounting, processing and paying, to Bank Reconciliations, to calculating MNL’s daily NAV, to creating Quant based Stock Universes etc etc.. We have built these Agents, tested them, adapted them and we have used a number of models in the process. We would stress that when building Agents, we have only used US Frontier Labs’ models with any success, as opposed to open weight models. We do use open weight models for inference through multi-model AI platforms like Perplexity. During this process, we believe we have learnt something that is very important in framing the AI investment story. However, to fully understand this key finding, I am afraid I am going to have to get into some detail of what we have done. Here goes.

This is a snip of part of our Agent Register:

Most of these agents do what they say on the tin. Most of our Agents utilise API connectors to access many other tools and data services, of which the most significant for us is the paid for, Bloomberg API (which is an additional subscription to the Terminal).

Just to digress again, what percentage of investment teams managing Funds in the UK do you think have this Bloomberg API linked into their chosen AI models? Surely some financial analysts think it is a good idea to ask this question to the people managing your money?

Let’s talk briefly about two agents. Both non investment, back office agents.

The Invoice Processing Agent. This agent receives a due invoice into its mail box (a##########@mlcapman.com) from the supplier. The agent extracts the relevant details from the invoice and posts them into a spreadsheet in Google Workspace (to which the Agent is connected). The Agent then works out what when and why we paid this supplier before and then posts a message to a human that looks like this:

Even better when the Human in the loop does not respond the Agent then sent a reminder and makes a piercing bell noise on their specific PC.

(Yes we have considered electrodes).

The payment is then Loaded by Human A, authorised by Human B and by Human C.

The agent then sends a Dash-Board Update such as below:

Note for the observant: If you look carefully, you can see one of the Humans has forgotten to Second Authorise a payment.

Across the M&M group we used to employ 5 in the accounts department. Now we employ zero. Nearly every accounting & financial function is agentified with Human in the loop only at sensitive points. We even have agents that complete all our FCA returns with very little to zero input from Humans apart from the checking process pre-submission. That is worth £ks just to not to have to listen to the goat noise from the previous human submitter.

I can categorically tell you several things with confidence: Agents work, Agents produce better results than Humans in the tasks discussed above, and they are cheaper. What is more relevant is that with the harness provided by the Labs, they are also controllable, auditable and accountable.

But there is a problem. To explain that problem I will give an example of a very simple use of AI. We have been going through the very tedious process of producing our Annual Report. The Company Secretary is responsible for compiling this document through the various changes and drafts. Each time we did a draft and received the next draft back, we would load the changes provided by ourselves, the auditors, the fund accountants, the Board into an Ai along with version x and version x+1 and ask the simple question: “Have all changes been made?” The response would be an excellent and instantaneous report that showed all changes not made, all changes made that were not asked for, all changes made incorrectly, and a list of further suggested changes. Done, in seconds. So we asked the Company Secretary why they could not undertake this exact same process to pick up the changes not made and changes made incorrectly, BEFORE sending a clearly erroneous version to us? The response was that they are not allowed to use any AI applications because their parent company forbids the use of AI. Yet they want their fees to increase every year with RPI! This anti-AI stance is utterly illogical as to use AI in this manner is a low hanging fruit productivity gain right in the core of their daily activities which is drafting documents and announcements. Yet, here we have Adoption Risk in the practical outturn of our operations.

BTW we like both the Humans at our CoSec, but if someone were to offer us an alternative AI at the core Company Secretarial service at a reasonable cost, we may switch. Interestingly, I was asked to remove that line from this Newsletter by a more sensitive member of my team. But I have not, for a simple reason: I very much doubt the UK has a Company Secretarial service provider with AI at their core.

Jensen Huang at the Goldman Sachs Communacopia + Technology Conference on 10 September 2026: “Everything’s coding. … When you codify a business process, that’s coding. … It’s also the easiest thing for AI to learn.”

Lets try to simplify and abstract what an Ai agent is.

An AI agent uses reasoning to decide how to achieve a goal, then orchestrates the use of tools and deterministic code to carry out the necessary tasks. We are all aware that Agents are exceptional at coding but the majority of AI commentators are awaiting its ability to diversify its skills into adjacencies such as finance, accounting, law, marketing etc.

Yet, Anthropic has consistently guided for enterprises to distinguish predefined workflows from Agents that dynamically decide their next steps, and recommends simpler workflows for well-defined tasks. We suspect that many AI commentators ignored this advice and hoped that the Agent would use its reasoning skills to work out the best way to undertake an enterprise function and then process that function also based on reasoning. We believe that latter approach is not suitable for this stage of the Era of Ai.

Why? The main reason is that effectively nearly every process in some vertical functions can be codified so why use non-deterministic reasoning when you can codify the process and then execute the function with an Agent with max_deterministic code which in turn reduces the probability of error. We use deterministic code wherever the rules can be specified, we are use Error Handling to catch edge cases and reserve model reasoning (under human supervision) and human judgement for the rare areas of ambiguity and exceptions not caught by our Error Handling.

So what are we saying? When we started building Agents, we also ignored Anthropic, and we would experiment with providing the agent a prose based narrative of the process that we wished it to undertake, gave it access to data it may wish to use and left it to do the job. The results were a series of failures that then adapted to a quasi-success. However, the agent could get confused easily and we were finding that it failed in say 10% of runs of a task, even after succeeding at 90% of runs of the same task, with very little differences in inputs.

“Pseudo-code is an informal, high-level description of an algorithm that uses human-readable language instead of strict programming syntax. It helps software engineers plan and map out the logical structure of a program before writing actual code.”

So, as people with degrees in Software Engineering, we changed our approach. We build the agent with pseudo_code, then we ask the frontier model to codify the process using as much deterministic code within a python script as possible. We even included Error Handling within this code, all written for us by the model. After all, coding is what AI is exceptional at doing. Let’s be clear we are effectively using AI to translate business knowledge of previously human undertaken processes and converting that into reusable software, then using that software to execute recurring work without a human.

Let me show you an example of a Skill.md file using pseudo code:

So the model then wrote the python code this Agent then runs, for example see Step5: “python scripts/update_nav.py”.

Now the more observant of you will have seen that python code in an image shown in an earlier section. I have shown it here again, look at the right hand side of the image. This is pure deterministic python code. Once validated, the same code can execute the specified rules repeatedly without requiring the model to reconstruct the calculation each time. It is highly reliable.

That is how we have got ALL our Agents to work in accounting, marketing, payments, invoice processing etc.

The snip above is an Agent that gets all the Asset Statements from our PBs via an FTP link and then uses these to calculate our NAV almost instantaneously. It’s not a simple job. Actually, the agent undertakes this process both at UK close and at US close and when we are presented our Weekly NAV calculation from our Fund Accountants the differences are normally miniscule.

This Agent works and does a deterministic job with accuracy and reliability twice every day at a cost of pence per day. Whereas, MNL’s Fund Accountant does a weekly NAV for us and costs the Fund over £200,000 per annum. The initial working version of this Agent took less than a day to build and test; validation continued through subsequent reconciliations. OK it is agreed that Fund Accounting services can include reconciliations, accounting policy reviews, reporting, audit support and other work such as Tax and Payroll, but you get my gist. Without daft regulation, we could get this Agent audited by a specialist, and we could save the Fund a significant share of it’s ongoing operating costs. It will happen one day.

So there are two glaring conclusions here: firstly, max_determinisitic agents work brilliantly, they cost peanuts and could save material amounts of money. As an enterpise, you would be insane not to agentify your basic day to day processes, as your competitors will. Secondly, can anyone with no Software Engineering skills build Agents like these? Yes they can if they understand systems well and can specify the process precisely, understand if the Model has codified the process correctly, validate the result and take responsibility for operating it. BUT will their employers allow them to implement these agents if the senior management are not engineers or systems specialists, and after how much testing and systems auditing?

This means there is another Adoption Risk. Building Agents that really work with complete confidence is not a “no code” or even “lo code” endeavour to use the industry parlance.

Portfolio implications for our findings

We also suspect that greater use of reusable, deterministic code could shift the balance of computation towards CPUs and reduce GPU inference requirements per task, as routine parsing, calculations and file processing can run without repeated calls to an AI model. Note our portfolio holds all of AMD, ARM and NVDA.

“There is nothing so useless as doing efficiently that which should not be done at all.” Peter Drucker

Whilst we are on this point, the lowest tier Fund Analyst keeps asking why we release Weekly NAVs rather than Daily NAVs. The staggeringly obvious answer to this confusion is that it would cost multifold more for very little added value. What do I mean by very little added value?

You can try this experiment at home at our next NAV date:

Open up your chosen AI application.
Upload either our latest Factsheet or the latest update portfolio that we regularly load on this site: https://www.linkedin.com/company/mnl-ln
Provide the date of the Portfolio Provided.
Provide the last published Weekly NAV and its date.

Then all you have to do is ask your chosen AI application to estimate the NAV right now/live and remind it to consider foreign exchange rates.

Every single time we have run this experiment the given NAV is within 0.25% of what we believe the live NAV is. By the way, all the market makers in MNL know they can undertake the above process too. We know that because we have educated each one of them. Whether they are doing it, is another matter. So, NO we are not going to burden our cost line with another £xxx,000 per annum to solve a problem that can be solved for zero marginal cost by anyone at home on a basic AI subscription plan.

We use the snip below to prove our point, check the number in the green highlight produced by our Agent and the number in green box produced by the Third Party Fund Accountant:

“You will never get to the end of the journey if you stop to shy a stone at every dog that barks.” Winston Churchill

On the Money Makers podcasts some weeks ago, Ben Rogoff said (17.23): “So they (agents) are as
good on the knowledge front as any human accountant, but they’re not brilliant at math. So what they can do now instead is
call up Python, run the maths in a sort of side pocket. This is quite technical, but I’ll get to the punch line. They’re
able to offload the cognition part of the workflow, but still be able to deliver the synthesis that no human can
match. And so, by being able to call tools like Python, you’ve turned something that is inherently
probabilistic into something increasingly deterministic. And now the tax return that you do in the morning
and the afternoon is the same number. And I think that’s one of the parts of this agentic story that maybe others.
have missed.

As ever, Rogoff is correct. He also said a few months back that the Technology is now proven, and the Return on Investment potential is now proven. So, as we stated earlier, we believe the challenges now lie in Regulation and Adoption.

P(extinction) = Nonsense

Our experience of building AI agents has made us increasingly sceptical of confident predictions about autonomous AI bringing about humanity’s extinction.

Why? Because the gap between impressive reasoning and reliable execution is substantial. When we asked agents to undertake simple enterprise processes using prose instructions and their own judgement, we encountered inconsistent results. Reliability improved when we used their exceptional coding abilities to turn clearly specified processes into tested, repeatable software.

The practical lesson is straightforward: use AI to help engineer the process, then minimise how much of its routine execution depends on the model improvising correctly.

That distinction matters for the extinction debate. Some scenarios appear to assume an agent capable of sustaining an extraordinarily complex undertaking, overcoming obstacles and coordinating actions without making the kinds of mistakes that routinely derail far simpler business tasks. Anyone assigning a probability to such an outcome should explain how the system crosses that reliability gap, over what timeframe, and with what evidence? Our findings show no evidence of such ability.

Writing code could help an agent overcome its unreliability. But it also creates opportunities for control. Code can be inspected and tested; its execution can be isolated; access to money, credentials and external systems can be restricted. Crucially, these controls can sit outside the model’s authority. The agent should not be able to grant itself new permissions or rewrite the rules governing its own execution. As Jensen says, its all a matter of good engineering.

None of this makes dangerous AI impossible. Deterministic means repeatable, not correct or benign. The same coding capability that makes a useful agent more dependable could make a harmful one more effective but this will require malicious intent driven by Humans. Future systems may also overcome the limitations we see today, especially through recursive self-improvement but we have seen absolutely no evidence of this ability from models so far.

We studied Reinforcement Learning techniques when undertaking our AI studies and the simple lesson is goals and rewards within RL have to be very precisely defined. Some of the experiments we have read about that have gone wrong look almost like Adversarial Testing, designed to fail by vagueness. If this is a surprise to some, it is not to ourselves. Our experience therefore challenges the confidence behind melo-dramatic extinction forecasts.

There is also a commercial question: who benefits from the proposed response to this melodrama?

Critics argue that restrictions and regulations shaped by the leading AI laboratories could protect those laboratories from competition. Licensing requirements, expensive evaluations and complex compliance obligations are easier for well-funded incumbents to absorb than for new entrants.

A coordinated slowdown could offer another benefit. If competitors were prevented from advancing as quickly, an incumbent might face less pressure to fund successive frontier training programmes. Its existing models could remain commercially valuable for longer, while reduced competition could ease downward pressure on prices. Under those conditions, a slowdown could improve margins through both lower competitive spending and stronger pricing power.

There is economic logic behind the accusation that these advocates are “talking their own book” with childish monster stories that raise questions about their credibility.

Utilised by human with bad intent, AI can be extremely malicious. Again, coming back to AI’s exceptional capabilities in coding, a massive cyber hacking threat IS approaching. Turning back to the UK Investment Trust industry, this is yet another question that a good financial analyst should be asking of Funds: “Can your Fund Accountant load payments into your Banking System? If so, are they storing all their passwords in a Password Manager application? Is the Password Manager application they are using USA, Chinese or Russian owned?” We would predict that within the next three years, we will see a number of listed UK companies/funds have their bank accounts hacked and funds misappropriated.

No MNL’s Fund Accountant can not load payments for MNL. They could, as they undertake payments for many of their fund clients, but we removed their permissions.

Zoomorphic language

We actually wrote about this in a Newsletter three years ago at the birth of the Era of AI, and warned that all sorts of scare stories would be driven into society’s psyche about the bad machine! However, we are increasingly struck by the human and biological language used to describe AI. Agents supposedly “want”, “scheme” and become “radicalised”, while groups of them become “swarms”, a term appearing in Dwarkesh Patel’s podcast coverage. These words carry associations far beyond their technical meaning. “Swarms” can conjure locusts or wasps or an overwhelming attack; “radicalised” evokes extremism, fanaticism and terrorism. The reader is invited to import familiar fears into an unfamiliar technology before examining what the software actually did.

History gives us ample reason to scrutinise this technique. Orwell warned that “if thought corrupts language, language can also corrupt thought”. He referenced the use of words such as “vermin” and “lice” used by the Nazis. Whereas the Stalinists have been keen on using the “dog” terminology, “these mad dogs of capitalism” as Andrei Vyshinsky ranted. Emotionally loaded descriptions can shape the conclusions people reach by shaping the picture they start with.

The use of such words is used as a weapon of deliberate manipulation, and it is a good signal to pause and ask what the words are persuading us to assume. Before accepting that an AI has become “radicalised” and wishes to “enslave humankind”, we should demand a precise account of its observable behaviour, its permissions and the controls that failed. We should assess the mechanism before inheriting the metaphor’s fear that the machine will enslave us and steal our women folk.

If you have noted the strange concept AI has on leg positions during Human Sex, so did we. We used this particular representation to re-illiustrate our point about the ludicrous claims on P(extinction).

Thomas Paine, The Age of Reason (1794): “All national institutions of state or religion, appear to me no other than human inventions, set up to terrify and enslave mankind, and monopolize power and profit.”

Anthropic

An important question for investors is whether the model that builds an agent needs to be the model that subsequently operates it. Often, it does not. A frontier model can design the workflow Agent, write the software and help test it. Once that process is established, much of the recurring work may be handled by a cheaper model, or by deterministic code requiring no model inference at all. On platforms such as Perplexity, selecting a different model can be as straightforward as changing a dropdown.

See below an Agent (we use for compiling and summarising analysts research on our stocks of interest), then note the drop down menu of models we can switch to use.

This matters because a substantial potential market for AI lies in undertaking existing enterprise tasks. Many invoice-processing, reconciliation, reporting and administrative steps have well-defined rules. They need reliable execution, rather than the most sophisticated model on every occasion. Indeed, Anthropic’s own engineering guidance recommends routing straightforward requests to cheaper models and reserving more capable models for difficult cases. If businesses increasingly adopt this approach, frontier laboratories may win the initial development work without capturing most of the subsequent processing volume (inference).

The bear case is therefore that the laboratories bear the enormous cost of advancing capabilities while competitors progressively commoditise yesterday’s breakthroughs. Stanford’s 2025 AI Index reported that the inference price for roughly GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024. That does not establish that every model’s token price must fall every year, but it illustrates the pressure on the price of a given capability. Meanwhile, the application provider (Perplexity in our case) may retain the customer relationship, proprietary data and workflow integration, leaving the model supplier competing for interchangeable processing work.

The Bull case argues that the relevant measure is the total cost of a successfully completed task, not simply the price per token. Enterprise work can be routine in its purpose yet difficult in its execution: incomplete information, contradictory documents, unfamiliar exceptions and changing circumstances all require judgement. A stronger model may justify its premium if it completes more tasks correctly, needs fewer retries or avoids expensive human review. Frontier models can therefore remain valuable throughout operations, including investigation, planning, exception handling and supervision, as well as in building and rebuilding the software. So the Bull case relies on the Reasoning abilities of the models, and as you know from the above, we question whether we are confident enough in the current Reasoning skills to be let loose on our own Enterprise.

Another Bull case is that the advanced models will be used to push out the productivity frontier. So for example if you are using a hypothetical $100 trillion Global economy, a 4% productivity improvement that translated fully into additional output would create $4 trillion of annual economic value. If the Frontier Labs can harvest 50% of that being $2trn on a margin of 40 per cent and put that on a multiple of 20 times you get a valuation of $16trn. If that is shared in a duopoly then that shows a future valuation for OAI and ANTH of $8trn each, a quadropoly then $4trn each.

A further extension of this point questions whether the real bull case may lie in enterprise services that scarcely exist today. Focusing exclusively on replacing accountants, programmers or customer service staff assumes that the market consists of a fixed inventory of existing work. But businesses currently leave countless questions unanswered and possibilities unexplored because investigating them costs too much. If AI dramatically reduces those costs, it could create demand for services that no company would previously have contemplated purchasing. Optimisations using Quantum and AI, immediately spring to mind. Consider services that continuously design and test bespoke materials against a manufacturer’s requirements, explore thousands of alternative product designs for individual customers, or simulate how combinations of suppliers, competitors and customers might respond to a company’s decisions. The opportunity would be to make experimentation and discovery available at a scale, frequency and degree of personalisation that was previously uneconomic.

Bulls also argue that there will be opportunities for the Frontier Labs to build specific applications for healthcare, manufacturing, scientific research and other valuable activities which are an attractive route to retaining more of the value created.

On a more flippant note, we are told the difference between Frontier and Open-weight offerings is the harness, yet I was working on the OAI platform (within harness) and this message was sent:

The extinction machine does not even seem to know where it is being used.

Our best guess on the best way to play the Frontier Labs is illustrated in this snip below:

Nearly every way you look at it (but there are efficiency cases where this does not work), if Open-weight wins then Hardware wins. If Frontier wins, then Hardware wins.

The investment returns of Ai are now proven

We mentioned this point above but we do wish to elaborate on this point more as we believe this is yesterday’s concern.

We could present many artefacts of proof such as the endorsement of MSFT of this slide prepared by Morgan Stanley:

…but we suggest you read the words of Andy Jassy, CEO of Amazon, below whilst remembering this is a CEO who was at the forefront of the Cloud Computing build out:

Remember, enterprises are still very early in using inference at scale in their current production applications. We long believed AWS could become a few $100 billion revenue business and now believe it will be at least double that and very possibly be $1 trillion annual revenue business for us in time with very appealing accompanying free cash flow and return on invested capital.”

“Earlier this year, we said we plan to invest approximately $200 billion in cash CapEx in 2026, the majority of which to support AI and AWS. At this level of spend and higher, we have clear line-of-sight to strong financial returns. I’ll explain why. There are two major parts of the investment, the data centers and the servers and networking equipment that go into them. These have different capital cycles. Data center capital is spent starting two years before we can put servers into them to start monetizing. Once a data center opens with servers plugged in, we start generating significant revenue right away and then get to monetize these data centers for 30-plus years without having to spend that start-up capital again. Servers and networking equipment operate on a shorter cycle. We typically purchase these a few months before putting them into service, so we have strong visibility into customer demand before we trigger the spend. If the demand isn’t there, we won’t spend the capital. For servers and networking equipment, on average, it takes a little less than three years to break even on that investment. The servers currently have a useful life of at least five to six years and most of our AI capacity these days is being contracted for at least five-year terms. That means that we’re driving significant free cash flow on the servers and network of equipment in the two to three years after we break even.” That as we get a few years out and the revenue growth outpaces the incremental CapEx growth, which will happen at some point, the resulting revenue, free cash flow and return on invested capital is very compelling. We’ve done this before in the first era of cloud computing just over a longer time horizon, where demand built more gradually than it has in AI.”

This is why we reiterate the Key concerns today are: Adoption, Power & Regulation, and why we have focused on Adoption in this Newsletter.

Key Tweets of the Month

Key Manchester & London Updates

Please remember that all Fund based news is now posted to: https://www.linkedin.com/company/mnl-ln
Twitter will only see posts on Ai, Technology and the Global Economy.