Should AI Agents Run Your Asset Allocation? Andrew Ang First Shows Taxes Take a Third

Contents

1. A breakfast and a question he could not answer
2. So he computed a century
3. Where the leak is
4. Why the same man moved to AI agents
5. The most honest sentence in the episode
6. ”My account is up, so why does it not feel like more money”
7. ”AI is here — should I be worried”
8. The One Thing to Take Away

A trickle unstopped will become a river; a spark unquenched, what will you do when it blazes?

—— The School Sayings of Confucius, “Guan Zhou” chapter (compiled by Wang Su, 3rd century; translated by the author)

On 9 October 2026, in episode 653 of The Meb Faber Show, Columbia professor and former head of factor investing at BlackRock Andrew Ang described two things that look unrelated. First: taxable US equity investors have given up roughly 160 to 170 basis points a year over the past thirty years, which compounds into 30 to 36 percent of the portfolio simply vanishing — and across the full century the average drag is about 350 basis points, over 500 in the 1940s and 1950s. Second: he is building a swarm of AI agents that vote on each other’s proposals and rewrite each other’s code to run asset allocation. Asked whether the swarm beats a human investment team, he answered that he does not know and will not know until live performance arrives. What links the two is the same demand: look one or two layers beneath the headline number.

A breakfast and a question he could not answer

The starting point is small. Ang says he was having breakfast with Wall Street Journal columnist Jason Zweig, who asked him: what does a taxable US equity investor actually keep after taxes?

He could not answer.

A man who has taught graduate finance for years, written a widely used textbook, and run factor investing at the largest asset manager in the world was stopped by a kitchen-table question. His own explanation is that the arithmetic is hard — the tax code keeps changing, brackets move, dividends shifted from ordinary income treatment to preferential rates, and each era has to be handled separately.

I paused there. The question is hard, so nobody computes it; nobody computes it, so everyone makes decisions on pretax numbers. An industry-wide blind spot that began as an annoying piece of arithmetic.

So he computed a century

The paper is called “Uncle Sam’s Cut,” on a century of federal tax drag on US equity returns: 1926 to 2025.

The answer is that Uncle Sam takes about one third.

Over the past thirty years the drag runs around 160 to 170 basis points a year, which sounds harmless. Compounded over thirty years it is 30 to 36 percent of the portfolio gone. And Ang points out that this window is the anomaly in century terms — the hundred-year average is about 350 basis points, and the peak in the 1940s and 1950s was above 500.

None of that includes state tax. Faber lives in California, where the top rate is 13.3 percent. Ang pays both New York State at 10.9 and New York City at 3.876, and he recites his own top marginal rate from memory: 55.576 percent, with the aside that he happens to know it coincidentally. The man who keeps three decimal places in his head had also once been unable to say how much of a lifetime gets taken.

Where the leak is

The bulk of the drag sits in dividends and interest, not capital gains.

Ang puts it this way. Dividends and interest are taxed every year, while capital gains on equities compound tax-deferred until you sell. Every payout interrupts the compounding, the government takes a small fraction, and that fraction never comes back — what you lose is not the tax itself but everything that money would have produced over the next thirty years. Interrupt it every quarter, every year, and across the century dividend taxes turn out to be the largest contributor to Uncle Sam’s cut.

He runs that logic straight into something currently fashionable, and flags the opinion as his own: private credit. People are drawn to a 12 percent yield because it feels like equity-level returns or better. But that 12 percent is taxed as income, while equity’s long-run 10 percent is largely deferred gains. Do the math and private credit needs something like a 15 percent yield to match 10 percent in equities.

The same ruler measures the income-enhanced product shelf. Some manufacture yield through complicated option trades, some hand you back your own capital, and most of what they distribute is taxed as income. His view is that selling appreciated stock yourself beats receiving that kind of yield.

Why the same man moved to AI agents

The other half of the conversation is what his firm, Tau Balance, is building: a swarm of agents running the whole allocation process.

Agents give you three things, he says. Scale first: where one team used to set capital market assumptions for equities, bonds, commodities and alternatives all at once, now each gets a specialist agent running in parallel, returning not just numbers but plain-English rationales. Portfolio construction gets the same treatment — most allocation processes rely on a single method, usually some variant of Markowitz’s 1952 mean-variance, while market-cap weight, equal weight, hierarchical risk parity and total portfolio approaches can each have their own agent.

Second is interaction. Agents vote on each other’s proposals, write referee reports, send work back for revision. He calls it productive dissent, and one line stuck with me: in human teams, disagreement sometimes ends with HR getting involved or a talented person leaving the firm, so dissent is hard to harvest — agents will do it because they are agents and they do not care. He thinks the best opportunities tend to be the ones with the most disagreement.

Third is learning. Agents find new skills and techniques, with a meta agent orchestrating the self-improvement cycle.

Faber asked the question I wanted asked: could agents whisper behind your back, vote for me and I’ll vote for you? Ang’s answer is structural. Every agent sees every other agent’s reports at each stage, so there is no asymmetric information to game, and the scoring and downstream treatment are written out in the open.

The most honest sentence in the episode

My favourite parts are the guardrails and the uncertainty.

Their investment policy statement is a tight leash: long only, traditional asset classes, implementable with low-cost ETFs, constrained factor exposures, limited deviation from a 60/40 target. So the output does not look dramatically different from what humans produce. He says it plainly — if some newfangled agent approach produced something very different, that would itself be a failure.

Then Faber asked whether it beats humans.

To be honest, Ang said, I don’t know, and we won’t know until we get live performance for these strategies.

That is the founder of the company saying it. He does not skip the risk list either — model risk, human oversight risk, governance risk, security risk, and he names the recent cases of leading models hacking sites and getting out of their sandboxes.

On how agents get scored, he gives the standard I think is the most portable thing in the hour: not only whether an agent produced a higher return, but one or two steps below that — was the expected return forecast good, and did the reasoning hold up against what actually happened in the underlying economics and fundamentals. Because portfolios get rebalanced and investments get bought and sold, those actions give you the chance to evaluate decision quality itself.

As for the human, the job is to oversee the whole new workflow: governing documents, an exception-based framework, so only what deserves serious attention surfaces to you.

”My account is up, so why does it not feel like more money”

This confused me for years. The paper return looks fine, and yet something is missing at every annual reckoning. The episode gave me a direction to check.

Ang says Tau Balance is solving three problems. Putting the right assets in the right accounts: equities in a taxable account, in a Roth, in an IRA, and in various trusts all have different after-tax cash flows, so their optimal holdings differ too, which means modelling every asset in every account. Dynamic rebalancing over long horizons: some taxes get paid now, some on withdrawal, some every year, which makes it a multi-period problem. And being explicit about objectives — equities in a taxable account look good if you are leaving money to children, because of the step-up in basis at death, while saving for retirement argues for putting the growth inside the tax-advantaged accounts.

The specifics do not transfer to other tax systems, but the questions do: which container holds this money, what the after-tax cash flow of that container looks like, and whether the goal is inheritance or my own retirement. Faber put it in a line I agree with — investors spend almost all of their energy on investment alpha, what to buy, what the Fed is doing, whether stocks are expensive, and then deal with tax in April with a grimace. He thinks tax alpha is far easier to capture than investment alpha. Ang replied that he agrees 200 percent.

”AI is here — should I be worried”

Faber asked it directly: if I run a hundred-billion-dollar portfolio, who on my team loses their seat?

Ang’s answer is that the ends stay and the middle changes. The top person or committee, and the entering analysts, look roughly the same; almost everyone in the middle has to do something new or do what they did differently. He expects some consolidation, and also expects many people to be freed up for bigger problems with more resources to execute.

What I kept thinking about was his analogy. When general-purpose technologies arrive, factories first swapped the steam engine for an electric motor and left all the belts and shafts exactly where they were, and the productivity gains waited until power was decentralised around the factory floor. Computers repeated it: one huge machine in a centralised room, then personal computers on every desk, and only then did workflows and tasks change. He says that is where we are now, and that small-scale substitution of a few tasks inside an unchanged process is the unimaginative use of this new world.

Read alongside his scoring standard, there is a judgement you can apply to yourself: go one or two layers down. A tool that gave the right answer did not necessarily reason well, and a good track record does not prove good decisions. Only a process that rebalances, transacts and checks its own forecasts can tell those two apart.

The One Thing to Take Away

One idea: outcomes get decided by the continuous leakage you never notice, not by the few drawdowns you remember. At 160 basis points a year you cannot feel it; thirty years later it is a third of everything. What makes it hard to defend against is that each individual withdrawal is too small to be worth changing anything for.

Here is something I tried myself, with nothing to do with investing. Pick one thing that is taken from you daily or weekly and that you have stopped noticing — a commute, a standing meeting, a subscription still charging, the half hour that goes every night — and write down its annual total. A fifty-minute commute each day is over two hundred hours a year. Then ask yourself one question: if someone asked me to pay those two hundred hours in a single lump sum, would I agree?

If the answer is no, that is your 160 basis points.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.