With all the discussion about AI’s progress this week, I keep coming back to how we govern what these systems actually do.
When an answer or action depends on an agreement, you should be able to see the rules behind it.
Here’s what I mean.
“Can I stop a prepayment?”
The answer might take you through three agreements. You should be able to follow the reasoning and see the language behind it.
If we use agreements to coordinate economic activity, why would we expect AI agents to operate without understanding them like this?
Understand · 1 of 4
If you work in digital banking, I'd love to hear how your bank is thinking about this.
JPMorgan, Citi, Bank of America, Wells Fargo and Morgan Stanley are blocking AI agents from customer accounts (American Banker, 27 Oct).
Jamie Dimon recently called letting just any outside agents in "just too risky for clients and for JPMorgan". He's not wrong. DAC's agent governance infrastructure would give banks a way to say yes.
Let's go over what happened. One large bank watched an AI agent reset a customer's password, change where money could be sent and move $100,000 out. Its fraud team went to the CEO and recommended blocking every outside AI agent, immediately.
But agents aren't going away. The real problem is that nobody can tell what an agent was actually allowed to do. Your bank sees a login, not your instructions. And if the agent goes off script, the fine print mostly puts it on you:
• Chase's terms: what an outside AI does in your account counts as yours, even if you didn't intend it
• Reg E, the federal transfer rule: hand someone your login to move money and you're "fully liable" for transfers beyond what you allowed, until you tell your bank
I think agents should be able to act for you, but only within what you actually agreed to, and in a way everyone can check.
So in the video we sketched a gate between your agent and your bank. It checks every request against the agreements behind it (your instructions, the agent's terms, Reg E and a new clause we're proposing for banks) and records every step, down to the moment you cut the agent off.
Say you tell your agent to pay the electric bill if it's under $200:
→ a $180 bill gets paid
→ a $240 bill comes to you first
→ a complaint you never asked for never goes out
1
2
142
If you lend into AI data centers, you'd want to know who pays when a tenant can't. On some deals it's a guarantee from Nvidia or Google, and those don't pay until the lease starts and stop at a cap.
And a lot of leases haven't started yet. Adding up the footnotes in the latest filings from Oracle, Microsoft, Amazon, Alphabet and Meta, I get about $1.3 trillion. Video's below.
Lately, per Bloomberg:
• 15 Oct: record cost to insure Oracle's and CoreWeave's debt
• 30 Oct: banks looking to insure their AI loans
But as of 7 Nov, I can't find a missed payment from any company in the video.
So this next part is a what-if.
Say more customers move to open-source models:
→ OpenAI and Anthropic earn less
→ most of what they owe stays the same, so they fall short
→ the guarantees pay, up to their caps
→ the rest goes to whoever holds the debt, like insurers and pension funds
Are you planning on buying the Anthropic IPO soon?
If so, let's take a look at what you inherit.
When you own a share, you own a slice of every agreement the company signed.
So I read through Anthropic's S-1 (filed 26 Oct) and its exhibits.
• Broadcom TPU leases: about $161B of rent before interest, on five-year terms. The lease calls it "absolute and unconditional"
• Cloud minimums: at least $114.5B with Google, $110B with Amazon and $44.9B with Microsoft, owed "regardless of actual usage" in the S-1's words
• SpaceX: up to $120.5B, and either side can exit from 2027 (reported)
Two lines in the S-1 add to $587B at 30 June (our sum). First-half revenue was $16.3B, and the S-1 calls it "primarily consumption-based". Fixed bills on one side, usage on the other.
Then there's Broadcom's side. It passed the racks and leases to two vehicles that investors fund, so the rent runs Anthropic → vehicles → their lenders. Broadcom backs part of that debt, up to about $29B on the first vehicle and about $43B on the second
So I ran a what-if. Suppose revenue grows slower than the commitments assume (hypothetical, not a forecast, not investment advice):
• the minimums would still be owed
• the rent would still be owed
• Anthropic could pay rent in notes, not cash, up to $42B, on conditions
• if the second vehicle's rent stopped, its racks could be sold
Say the racks fetched 50 per $100 owed (illustrative). Broadcom's filed formula would pay the gap to 72.5, so the lenders paid first would be covered up to the cap. The second-priority lenders would be paid beyond that.
You own part of the company owing the rent and the minimums, and the agreements decide who gets paid and in what order. It's all filed, but it's spread across the S-1, its exhibits, and Broadcom's own filings, and you only see it when you read them side by side.
1
138
Hey Lenders! What if your borrower blames AI?
Bloomberg reported on 27 Oct that force majeure ("act of God") clauses are appearing in credit agreements for buyout loans. If the borrower says AI caused a missed payment, the lenders' remedies pause. Interest is still owed, but there's no clear end date.
Here's an actual transaction over the past 14 days:
H.I.G. got it into a €470M loan for Terras Group, which builds rail and foundations, not software.
However! Lenders had similar language struck from Bain's €4.7B-equivalent Everllence loan.
Normally the path after a missed payment is written into the loan on day one:
missed payment → a few business days of grace → remedies open → lenders holding two-thirds can make the whole loan due → security enforced.
So I ran a what-if on an illustrative loan (hypothetical, not investment advice). With the clause in:
• the grace would last as long as the borrower felt AI's impact
• the borrower would decide it was AI, as reported
• calling the loan and enforcing security would be on hold
But the clause only pauses the lenders. Nobody else:
• a CLO holding the loan would still mark it down, grace period or not (one filed CLO indenture as the example)
• if it's rated, S&P would put it at 'D', its lowest grade, within 30 days, however long the grace runs
The remedy would pause, but the consequences... yeah, they don't.
That's what we built DAC for. It reads the loan before you sign it and shows what each clause does, to you and to everyone downstream.
Lenders/ credit X, would you sign this clause? Yes or no?
2
117
There's so much talk right now about AI and debt, but most of it focuses on using debt to fund AI. I keep thinking about the other side: AI actually reading the debt...
Every one of these deals runs on a credit agreement/ bond indenture, and AI can already read one and return a correct, cited answer.
The trouble starts when you ask again.
Ask whether a borrower can license its IP to a subsidiary and one desk hears "restricted," another hears "yes, with conditions," and a third gets "needs review." You can use the same model with the same agreement and get three different answers, because the AI rebuilds its reading every time you ask.
So how do you fix that?
You give the AI something solid to work from. Map the agreement's rules once, with the exceptions and sources attached, and have the AI work from that instead of starting from scratch every time someone asks. Then everyone gets the same answer, with the conditions behind it. And when the facts change, you just run them through the same rules.
Same models, different system. That's what this 👇 video is about.
💯
Knowledge graphs have always felt like the holy grail of legal analysis. Huge potential, but surprisingly fragile and difficult to get right.
I've been graphing patents since mid-2024, initially in Python and more recently in Neo4j/Cypher.
One of the most difficult aspects is figuring out how to traverse the graph to retrieve the right context for a particular analysis. I.e., GraphRAG.
Below is a graph showing relationships between patents, with different patent families grouped by color.
1
109
DRM 2.0 incoming
... Just a matter of time
REA (Reverse Engineer Anything) by @morluto is now a skill in Computer.
Ask Computer to reverse-engineer an app and it inspects how the app works without its source code, then builds a working version.
Tasks that use /reverse-engineer are free through October 31, up to 10,000 credits per user.
Here are 5 apps Computer made with it, available to try for a limited time:
1) Figma: Design with layers and auto-layout, plus sign-in, cross-device sync, and multiplayer editing. Try it: ui-editing.pplx.app
1
146
Just look at how these animations are built and loaded. It reminds me of the dial-up era (on the left) while transitioning to cable modem/DSL/T3 (on the right)
Now think about real-time inference for robots. We're just getting started...
Does your agent know what you agreed to?
An AI agent can stay inside its guardrails and still pay $3,000 for a report nobody checked
Why it matters: agreements are the operating system of economic activity. Humans and institutions have been the runtime. Now agents are joining it
Those agents are getting guardrails to stay in bounds, and specs to find each other, check out and pay. But what two companies agreed to is in the agreement, and the easy way to wire a job leaves it out:
provider's agent says "complete" → payment fires
Nobody verified the report and the buyer never accepted it. The agreement was never asked
So how do you keep the agreement in charge? The payment waits for it:
report delivered → verified → buyer accepts → paid
And every action gets a gate that pulls the rules from the agreements and checks them before anything moves:
- Accepted, the terms are met
- Requires approval from a named person
- Blocked, the agreement prohibits it
That's what DAC is, an agreement layer for the agentic economy. The legal agreement stays as is, we build the machine-usable rule layer underneath it.
You and your agent should know the options before either of you needs them.









