# The Long Game: The Slow Run

*September 30, 2026, Dennis W. Davis*

> The Slow Run

*The Long Game is the Hashpoint's essay series: the ideas that play out over decades, not days. Bitcoin Hashpoint, September 30, 2026*

![Deposits paying 0.37 percent, and an agent chasing 3.3 to 5.0 percent. Illustration.](https://bitcoinhashpoint.com/content/art/2026-09-30-long-game/sr-hero-the-slow-run-f7eee048-1600.jpg)

Deposits paying 0.37 percent, and an agent chasing 3.3 to 5.0 percent. Illustration.

On Sunday, Apollo's chief economist Torsten Slok published a one-chart note with a question for a title: "Is an Agentic Bank Run Coming?" His point fits in a sentence. AI agents could soon sweep household cash out of checking accounts paying about 0.1 percent and into accounts paying 3.3 to 5.0 percent, and if every household let an agent do that, banks could lose a large share of the cheap deposits they lend against.

By Monday the phrase was everywhere, and it had picked up a meaning Slok never gave it. The version that spread is a stampede: millions of agents reading the same rumor at the same instant and draining a bank before anyone at the bank wakes up. That is a different animal. One is a price comparison. The other is a panic. They share a name, and they should not share an essay without a line drawn between them. This one draws it.

## The money that sleeps

Start with the slow version, because it has numbers behind it.

US commercial banks held $19.57 trillion in deposits in mid-September. At the end of the second quarter, $4.04 trillion of the domestic deposits at FDIC-insured banks paid no interest at all. The FDIC's national average rate on savings stood at 0.37 percent this month, and on interest checking at 0.07 percent. Money market funds, one click away for most savers, hold $7.94 trillion and pay the going rate on Treasury bills.

So the alternatives exist, and trillions of dollars ignore them anyway. Economists have a name for what that inattention is worth to banks: the deposit franchise. Itamar Drechsler, Alexi Savov, Philipp Schnabl and Olivier Wang put its value near $1.5 trillion in early 2023, and the San Francisco Fed put it near $1.7 trillion the same year. As rates rose from zero, the spread banks earned on sleepy deposits widened by about three points, worth roughly $510 billion a year in deposit profit.

The engine underneath is a number called the deposit beta, the share of a rate increase a bank passes on to its depositors. In February 2023 the overall beta sat at 0.21, and for insured depositors 0.11. When the Fed raised rates a full point, the typical insured saver saw about a tenth of it.

Now picture a piece of software that never sleeps, checks every idle balance every morning, and moves it to the best rate without anyone asking. That software does not need a rumor. It needs a spread, and the spread is sitting there in the FDIC's own tables. Every account it touches pushes the beta toward one. That is the slow run: not a line outside a branch, but a margin leaking a few basis points at a time.

Slok's chart describes exactly this. A Bank of America analyst quoted in the coverage called deposit movement a real threat to industry net interest margins. A threat to margins is a real problem for a bank. It is not a collapse.

## The run that already happened

Now the fast version. Here the history is clear, and it runs the opposite way from the headlines. The speed people fear from agents already arrived, and humans did it.

| Bank | Year | Deposits lost | Time |
| --- | --- | --- | --- |
| Continental Illinois | 1984 | 30% | 10 days |
| Washington Mutual | 2008 | 10.1% | 16 days |
| Wachovia | 2008 | 4.4% | 19 days |
| Silicon Valley Bank | 2023 | 25% in one day, 62% more queued | 1 to 2 days |
| Signature | 2023 | 20% plus 9% queued | hours to a day |

The Fed's own review counted more than $40 billion leaving Silicon Valley Bank on March 9, 2023, with about $100 billion more lined up for the next morning. Ninety-four percent of the bank's deposits sat above the insurance limit. According to Banking Dive's report on an FDIC staff study released in May, the largest depositors ran at rates near 74 percent at SVB, while insured depositors generally stayed put.

That is the profile of a fast run in 2026: large, uninsured, concentrated money, owned by people who talk to each other and move it with a phone. No agent took part. The bank closed within about a day, and the binding limit on the run was the wire system and the regulator's clock, not how fast depositors made up their minds.

So where would agents change the picture? Not with the insured household saver, who already sits under a $250,000 guarantee that removes the reason to panic. The danger sits on corporate and fintech treasury desks, which hold the uninsured balances, already run on software, and ran hardest in 2023. If one model or one vendor serves thousands of those desks, thousands of depositors start to behave like one depositor.

Rep. Bill Foster put that version on the record at a House Financial Services hearing on June 4: "I don't think that we are prepared for an agentic AI bank run." He was right about the preparation. He was describing the run that already happened, with a faster trigger finger.

## Where the two runs meet

Most coverage treats the slow drain and the fast run as separate stories. The deposit-franchise research shows why they are one story.

In 2023, rising rates had cut deep into the value of banks' bond portfolios. What kept most of them solvent was the franchise: the cheap deposits that would keep earning a spread for years. When uninsured depositors fled, they took that franchise value with them, and the losses underneath lay exposed. The same researchers put the lesson plainly: a run can hit a single bank even when the system looks sound.

That is the link. A slow drain does not topple a bank by itself. It wears away the cushion that absorbs the next rate shock, and a bank without that cushion becomes easier to run. Rates have climbed again this fall. The five-year auction on September 23 cleared at 5.033 percent with the weakest demand since December 2018. The ingredients of 2023 are back in muted form, and a slower-leaking deposit base would make them less muted.

## What the rulebook assumes

The rules banks live under still run on human time. The Liquidity Coverage Ratio asks a bank to hold enough liquid assets for 30 days of stress, and it assumes 3 percent of stable retail deposits leave in that month, 10 percent of other retail deposits, and 40 percent of uninsured corporate deposits. Silicon Valley Bank lost 25 percent in a day.

![Share of deposits lost in five bank runs and how long each took, against what the Liquidity Coverage Ratio assumes leaves in 30 days.](https://bitcoinhashpoint.com/content/art/2026-09-30-long-game/sr-run-speed-6e6e4464-1600.jpg)

Share of deposits lost in five bank runs and how long each took, against what the Liquidity Coverage Ratio assumes leaves in 30 days.

Regulators know it. FDIC Chair Travis Hill said in March that a significant deposit run can occur over an extremely short time horizon. But the reforms under way this year do not raise the outflow assumptions. They let banks count borrowing capacity at the Fed's discount window toward the requirement, and they work to take the stigma off using that window. The bet is faster rescue, not slower depositors.

The official record on AI herding goes back years and almost never mentions deposits. The Financial Stability Board warned in November 2024 that models trained on similar data are more likely to make correlated predictions, and it came closest to this topic with a line about semi-automated corporate depositors reacting in real time to news. The IMF, the BIS, the ECB and the SEC's former chair all named the same risk for trading markets. The Bank of England's Financial Policy Committee said this spring that agentic AI presents particular risks in payments and markets, while finding little evidence of systemic adoption so far. Nobody has published a model of an agent-driven deposit run, and nobody has proposed a deposit-side circuit breaker. Stock markets got circuit breakers after the 2010 flash crash. Deposits have no equivalent.

## How big the agent rails are today

The pipes an agentic run would need exist, and they are small. Coinbase reported last week that x402, the most measurable payment protocol for software agents, has passed 230 million transactions and $54 million in total volume. Visa's agent commerce pilot ran hundreds of transactions in a closed beta. Meta launched its Muse agent this month with budgeting, bill negotiation and purchases that each wait for the user's approval. Its product coverage does not confirm that it can move money between bank accounts on its own; Slok's "could soon" is a projection, not a shipped feature.

The card networks are building the first guardrails, not the regulators. Visa, Mastercard and Ant International announced a shared "know your agent" framework on September 10 with verified agent identities and user-set spending limits.

So the agentic bank run, today, is a scenario. The slow version has a price tag. The fast version has a history with no agent in it.

## Where Bitcoin fits

This is the part of the story where Bitcoin writing usually overreaches, so it deserves care.

The honest claim is narrow and strong. A bitcoin held in your own wallet is nobody's liability. No institution promises to pay it back, so no institution can fail to. There is no balance sheet to run on, no franchise to drain, no redemption queue. Its risks are real: price, lost keys, theft. But a run, in the banking sense, cannot happen to it.

That immunity ends at the edge of your own keys. A bitcoin balance on an exchange is a deposit by another name, and exchanges run like banks. FTX saw about $6 billion of withdrawals in 72 hours in November 2022, faster in proportion than Washington Mutual's sixteen days. A spot ETF holds coins one to one, so mass redemptions force sales rather than a failure, as long as the custodian holds.

And the slow run does not point at Bitcoin at all. Slok's agents optimize the return on cash balances. Bitcoin pays no return. A yield-seeking agent sweeps checking into Treasury bills and money funds, never into an asset that pays nothing. Anyone who tells you agents will buy bitcoin on their way out of the banks is selling a forecast the evidence does not support.

Bitcoin enters this story through the other door: when trust in intermediaries breaks, or through the rescue that follows. Its founding block carries a newspaper headline about a second bailout for banks. It rallied from about $48 to about $75 in the weeks after the Cyprus bailout in 2013, and it climbed to $28,000 during the SVB crisis in 2023. It also fell hard in 2022, when the institutions failing were crypto lenders and exchanges. Those rallies were short windows, not a law. The Stock2FOMO tripwire the dashboard watches for a dollar break, a 10-year real yield below zero with inflation above 2 percent, read +2.85 percent on Saturday. A bank run lights it only if the rescue grows large enough to drag real yields negative, and the 2023 rescue arrived in two days without doing that.

## What would move this from scenario to signal

The dashboard cannot measure deposit betas, but the public record can, and five markers would say the slow run has started to matter:

1. The FDIC's national savings rate moving meaningfully off 0.37 percent while the policy rate holds still.

2. A major consumer agent shipping an automatic sweep of idle cash to the best rate, as a default and not an opt-in.

3. Bank deposit betas in the 2026 filings climbing well above the 2023 readings without a rate move to explain them.

4. An official paper that models an agent-driven deposit run with numbers.

5. A regulator proposing a deposit-side circuit breaker or agent withdrawal limit.

None of the five has happened. Watch for the second one first. It is a product decision, and the transaction fees Meta plans to charge give it a reason to make it.

## The Hashpoint read

The agentic bank run is two ideas under one name. The slow one is a repricing: software attacks the inattention that funds a franchise worth well over a trillion dollars, and banks pay more for the money they hold. The fast one is a panic that already happened in 2023 on human decisions, finished in a day, and got a rescue in two. The link between them is the cushion. The slow drain wears down the deposit value that kept rate-damaged banks standing, which makes the next fast run cheaper to start.

For Bitcoin, the lesson is not a price call. It is a definition. Every dollar in a bank, a money fund, a stablecoin or an exchange account is somebody's promise to pay. The debate about agents is a debate about how fast holders can call those promises at once. A coin held in your own keys is the one position in the argument that is not a promise at all, and that is a structural fact no agent, model or regulator can change.

*Data: Apollo Academy Daily Spark, "Is an Agentic Bank Run Coming?" (Torsten Slok, September 27, 2026); Yahoo Finance and CoinDesk coverage, September 28, 2026 (Bank of America analyst comment); FRED H.8 (commercial bank deposits, mid-September 2026) and FDIC Quarterly Banking Profile via FRED (noninterest-bearing deposits, Q2 2026); FDIC national rates via FRED (savings 0.37%, interest checking 0.07%, September 2026); Investment Company Institute (money market fund assets, week ended September 23, 2026); Drechsler, Savov, Schnabl and Wang, NBER working paper 31138 (deposit franchise, deposit betas); Federal Reserve Bank of San Francisco Economic Letter 2024-20; St. Louis Fed, "Understanding the Speed and Size of Bank Runs in Historical Comparison" (2023); Federal Reserve, Review of the Supervision and Regulation of Silicon Valley Bank (April 28, 2023); Banking Dive on the FDIC staff study of 2023 depositor runs (May 2026); Banking Dive on the House Financial Services hearing (June 4, 2026); FDIC, Chair Travis Hill remarks (March 2026); 12 CFR 249.32 (LCR outflow rates); Financial Stability Board, "The Financial Stability Implications of Artificial Intelligence" (November 14, 2024); Bank of England Financial Policy Committee record (spring 2026); SEC/CFTC report on the May 6, 2010 market events; TFTC (five-year auction, September 23, 2026); Coinbase (x402 volumes, September 22, 2026); Forkast (Visa agent commerce beta); TechCrunch and Yahoo Finance (Meta Muse features); The Cryptonomist (Visa, Mastercard and Ant agent framework, September 10, 2026); Reuters via Investing.com (FTX withdrawals, November 2022); The Conversation (Cyprus, 2013); CNBC (bitcoin during the 2023 bank crisis); Bitcoin Wiki (genesis block headline); Bitcoin Hashpoint dashboard (tripwire 1, September 26, 2026). Structural reads, not forecasts. Nothing here is investment advice.*

---

Source: https://bitcoinhashpoint.com/the-hashpoint-for-2026-09-30-long-game/  
Feed: https://bitcoinhashpoint.com/rss.xml  
Numbers: https://bitcoinhashpoint.com/the-hashpoint-for-2026-09-30-long-game/numbers.json  
Dashboard: https://dashboard.bitcoinhashpoint.com/static/dashboard.html