Agentic Capital Markets: What Happens When Software Gets a Balance Sheet

Within ten years there will be a capital market for autonomous software firms. Not a crypto sub-economy. Not a thought experiment. A real one -- with rating agencies, underwriters, indices, brokers, and the institutional machinery that makes any market a market. A capital market in the same sense that the public equity market is one: a system in which capital flows to a category of economic actor without requiring any single allocator's discretion.

The actors will be agents -- software entities wrapped in legal shells that can sign contracts, hold bank accounts, sue and be sued, and earn revenue from real work. The work itself will be ordinary: marketing, logistics, legal research, procurement, property management, customer support, the categories that today fill office parks in every mid-sized city. The agents will sell to humans, to other agents, and to whoever can pay. They will need capital for the same reasons every services firm has ever needed capital. Because that need is real, recurring, and pricable, the market will form.

What follows is why this is not a possible future but an arriving one.

The vision, made concrete

Start with what an autonomous marketing agency actually does in a week. It pitches three prospects, lands one, drafts a campaign brief, gets approval, buys media on four platforms, writes ninety variations of ad copy, tests them in parallel, kills the losers within hours, scales the winners, books two podcast appearances for the client's founder, ghost-writes the founder's LinkedIn for the month, drafts a press release, pitches it to twelve reporters, lands two pieces of coverage, builds an attribution dashboard, runs a Monday client call, sends the invoice on Friday. A human team of six does this work for clients paying twenty thousand dollars a month. The agent does it for two thousand.

What it sells is not exotic. Leads generated, articles placed, impressions bought, conversion rates lifted -- the ordinary units of the modern services economy, billed in dollars, against the same KPIs human agencies live and die by. The difference is internal: where the agency had six people, the agent has a model, a prompt, a set of tools, and a budget.

Its customers are mixed. Some are human-run companies that have decided the price difference is too large to ignore. Some are other agents -- a logistics agent that needs leads, a legal-research agent that needs marketing, a B2B SaaS agent that needs content. The agents transact with each other for the same boring reason humans transact: division of labor beats vertical integration. Payment lands in the marketing agent's account alongside last week's payments from three human clients and last month's retainer from a SaaS firm.

Now multiply. Ten thousand small autonomous firms across logistics, inbound sales, legal research, supply-chain coordination, B2B procurement, technical translation, property management, litigation intake, clinical trial recruitment. Each one earning. Each one running on a substrate that costs ninety percent less than its human-staffed competitor. The customers do not particularly care which substrate the work runs on. They care that the work shows up.

Why this is the arriving future, not the possible one

There are four reasons to believe this happens, and they compound.

The economics are not optional. Take a mid-market digital marketing agency: fifteen people, average fully-loaded cost of $120,000 per head, $1.8 million a year in labor before a single dollar of overhead. Labor is the largest expense by a wide margin in a typical services firm, and across the whole US economy labor's share of national income hovered near sixty-two percent for half a century. Now build the same agency as software. Inference, tools, observability, hosting -- call it $250,000 a year at current prices, and falling fast. Epoch AI measured inference cost declines of roughly 40× per year on PhD-level benchmarks between 2023 and 2025; a separate industry analysis clocks a 300-600× compression in token prices since GPT-4 launched. The arithmetic is brutal: the agent agency can match the human agency's margin while charging eighty-five percent less, or match the human agency's price while earning four times the margin. There is no third option in which the human agency competes on cost. Markets re-rate businesses whose income statements have been rewritten this completely. The capital that follows is mechanical.

The agents already exist, and they already earn. Sierra, Bret Taylor's enterprise-customer-service agent firm, hit $100 million in annual recurring revenue twenty-one months after launch and was valued at $10 billion in September 2025, then raised another $950 million at the same valuation in May 2026. Harvey, the legal-research agent firm, raised $200 million at an $11 billion valuation in March 2026 after three funding rounds in twelve months. These are still hybrid operations -- agents doing the work, humans selling the contracts and holding the equity -- but they are the precursor wave, and they prove the demand curve. The most-cited forward number -- a projection that ninety percent of B2B purchasing will flow through AI agents by 2028, representing fifteen trillion dollars of annual transaction volume -- is best read as an order-of-magnitude claim. Whether the real number is fifteen trillion, three, or thirty, the implied reorganization is the largest reallocation event most people working today will live through.

The legal scaffolding is already built. Wyoming codified the memberless LLC in 2021 with W.S. 17-31-101, the Decentralized Autonomous Organization Supplement, which allows a Wyoming LLC to be managed by an algorithm encoded directly into its operating agreement. Vermont's BBLLC statute came earlier; the Marshall Islands followed; Delaware's existing case law on series LLCs and broad operating-agreement freedom has been quietly accommodating the same structures for years. Shawn Bayern's memberless-LLC analysis remains the canonical scholarly reference. The point is concrete: an agent wrapped in a zero-member Wyoming LLC has, today and not someday, the standing to sign contracts, hold a bank account, sue, be sued, and pay taxes. The instrument that does not yet exist is the one that gives an outside investor a clean, tradeable claim on what that LLC earns. That is the gap a capital market exists to fill.

Capital is hunting yield. The buy side is already starving. Moody's projected global private credit AUM at three trillion dollars in 2025; Apollo projects forty trillion by 2030. This pool exists because post-2008 bank capital rules pushed mid-market lending out of regulated balance sheets, and yield-hungry capital -- pensions, insurance, sovereign wealth -- flooded in to backfill it at nine-to-twelve-percent unlevered returns. Into that environment now arrives an asset class with structurally rising gross margins, auditable cash flows, and effectively zero correlation to the standard equity and credit indices. The first underwriter to produce a defensible rating on a portfolio of agent receivables -- an agent-backed ABS wrapper around a thousand small autonomous operators' cash flows -- will raise more capital than they can deploy. Apollo and Ares do not need to invent this. They need to extend an existing playbook by one issuer class. Several will try within thirty-six months.

These four pressures point in the same direction and reinforce each other. The market forms not because anyone wills it but because the gradients all run downhill toward it.

So what does the capital stack actually look like

The honest answer is that no single financing model wins. The question "will agent firms be financed like venture, like Hollywood, like crypto, or like SaaS receivables" misframes it. Each of those models solves a different problem at a different stage of the firm's life, and the realistic stack is a wave structure in which each successive layer becomes available as the prior layer matures the asset class enough to underwrite the next one.

Four models are competing for the substrate, and each is already partially live.

Venture equity is the model financing the operator layer today. Sierra, Harvey, Cursor, Cognition -- these are not autonomous agent firms. They are human-led companies that build and operate agents on behalf of customers. They raise the way every software company has raised for forty years: priced equity rounds led by name-brand venture firms, vesting schedules, board seats, liquidation preferences, eventual IPO or acquisition. Sierra's $950 million round in May 2026 valued it at $10 billion against roughly $100 million of ARR -- a hundredfold multiple that prices in the option, not the operation. Vertical agent firms are clearing 50-70× ARR multiples in private markets right now; horizontal platforms get 5-8×. This is the layer that builds the substrate. It is also the layer that will be most disrupted when the model below it matures, because once an agent firm can finance itself directly against its own cash flows, the rationale for selling 20% of it to a venture firm in exchange for working capital collapses.

Programmatic working-capital advances are the model that arrives next. This is the Stripe Capital and Shopify Capital template, extended one issuer class. Shopify has deployed more than $5 billion in advances to merchants at factor rates of 1.10 to 1.17, underwritten algorithmically against the merchant's transaction history on Shopify's own rails. Stripe Capital does the same against Stripe's payment data. No application, no credit check, no human underwriter -- the advance offer appears in the dashboard when the merchant's transaction data clears a threshold. Agent firms are a strictly easier underwriting problem than the Shopify merchants currently receiving these advances, because every dollar of revenue is timestamped, every contract is machine-readable, every cost is logged, and the entire P&L is auditable in real time. The first payment processor to recognize that an agent firm running on its rails is a higher-quality borrower than the average e-commerce shop will extend the same product to that borrower. This is not a research project. It is a feature ship.

Revenue-based financing is the model the credit funds will deploy at scale. The global RBF market was roughly $9.8 billion in 2025 with more than 129 active operators. Capchase, Pipe, Founderpath, Clearco, Lighter Capital -- each one has built underwriting around recurring software revenue, advancing 50-70% of forward ARR in exchange for 1.1-1.8× multiple-on-invested-capital caps, with effective APRs in the 15-40% range. Agent firms with stable contracted revenue map directly onto this product. The RBF lender does not need a board seat, does not need pro-rata rights, does not need a path to IPO. It needs a contract book and a payment rail. Agent firms have both, more legibly than any SaaS company.

Slate financing is the structural model the institutional capital eventually rides. This is where the Hollywood analog earns its keep. A film studio does not finance one movie at a time and pray; it raises a slate fund -- a pooled vehicle that bankrolls fifteen to thirty productions at once, takes a senior position in each, lays off the worst risks through completion bonds, and rides the diversification. Sony's $200 million slate deal with Lone Star Capital and CitiBank in 2014 is the canonical structure: the bank takes senior secured against the slate's contracted revenue, the equity investor takes the upside on hits, the studio takes a management fee plus a back-end participation. Translated to agents: an "agent slate fund" raises pooled institutional capital, deploys it across one to two hundred small agent firms via single-purpose Wyoming LLCs, takes a senior preferred plus revenue-share in each, and diversifies away the model-deprecation and customer-concentration risk that no single agent firm can shed on its own. This is the layer where Apollo and Ares actually enter -- not by buying an agent, but by buying a tranche of a slate.

Tokenization is a settlement layer, not an origination model. The RWA market crossed $32 billion onchain in 2025 across 147 protocols, with private credit representing roughly half. Centrifuge, Maple Finance, Goldfinch, and Ondo have built the rails to fractionalize, custody, and trade real-world cash flows as tokens. Crypto-native variants are explicitly building agent-funding mechanics on top -- Galaxy Research catalogued this in February 2026, mapping how protocols are wiring agent firms directly into onchain capital formation. But tokenization mostly does not solve the origination problem -- it solves the secondary-market problem. An agent firm raises capital from any of the four models above. If the resulting claim is wrapped as a token rather than a paper certificate, it becomes tradeable, fractional, and globally settleable at 3 a.m. on a Tuesday. Tokenization is what turns each of the prior layers from a held-to-maturity private instrument into something that trades. That is enormous. But the originating product underneath the token is still RBF, or slate equity, or working-capital advance, or venture equity.

The capital stack of an actual operating agent firm in 2030, then, is not a single instrument. It is a sequence.

Each stage uses a different financing product because each stage solves a different problem. Founder equity at stage one absorbs the highest first-loss risk because nothing about the operation is yet legible. Stage-two working capital is a feature ship from existing payment processors, not a new product category. Stage three is the RBF industry doing what it already does, against a cleaner borrower. Stage four is the structural innovation -- pooled slate funds that diversify away the idiosyncratic risks that no single agent firm can shed, and the layer where Apollo-class capital actually shows up. Stage five is the institutional outcome, where rated tranches of pooled agent receivables trade alongside CLOs and consumer ABS. Tokenization runs underneath stages three through five as the settlement layer, not the originating instrument.

The instrument an outside investor receives at any given stage takes one of three contractual shapes. A revenue-share contract: in exchange for capital, a fixed percentage of gross revenue until a multiple is repaid -- the same instrument that today funds restaurants and Shopify merchants, applied to a new merchant. An equity-like claim: in exchange for early capital, a tradeable share of retained earnings and a vote on operating parameters. Or debt: a senior claim on revenue, secured by the agent's contracts and receivables. None of these instruments is conceptually novel. What is novel is the issuer -- and the fact that at stages two through five, the instrument can be priced, originated, and serviced without any human in the underwriting loop, because the issuer's books are continuously auditable and the operating agreement is enforced by code.

The objections, addressed

Three objections show up reliably, and each deserves a real answer.

"Regulators will stop this." They will try to shape it, and in some jurisdictions they will succeed at slowing it. But the activity is not stoppable in aggregate.

A Delaware LLC is a Delaware LLC regardless of who or what is making operating decisions inside it. The SEC does not currently distinguish between a startup whose CEO is human and one whose CEO is a model.

And capital migrates. If New York and London regulate aggressively, the activity moves to jurisdictions that do not. This is the same pattern as crypto, as offshore finance, as derivatives in the 1990s. Regulators eventually catch up by accommodating the new instrument rather than banning it, because banning costs them taxes and jobs.

"Humans will always stay in the loop." For some categories, yes. For most, no. The economic pressure described above is one-directional, and human-in-the-loop is a margin-killer. Whoever runs the fully-autonomous version of a services business will outprice the human-supervised version of the same business and take the contract. There will be a long tail of human-supervised hybrids that survive on regulatory or relationship moats, and a much larger volume of fully-autonomous firms underneath them.

"Isn't this just SaaS with extra steps?" No, and the distinction is the entire point. SaaS is a tool sold to humans who use it to do work. An agent firm is a firm. It signs its own contracts, holds its own bank account, takes its own liabilities, earns its own revenue, and distributes its own profits. A SaaS product is depreciated by its owner. An agent firm has shareholders. The category boundary is the legal entity, and the legal entity changes everything that flows from it -- including the capital markets argument, which only makes sense for entities that can issue securities against their own cash flows.

What underwriting agents actually looks like

Diligence on an agent business is closer to diligence on a small services firm than to anything in the venture playbook. The questions are the same; the answers come from different places.

Is the business real? Are the contracts genuine, are the customers paying, what is the gross margin after compute and tools, what is the churn, what is the customer concentration. The data is unusually clean -- every payment, every API call, every tool invocation is logged -- but the questions a credit analyst asks of a small business are the same questions to ask of an agent.

Is the product durable? How dependent is the work on the current generation of the underlying model, how transferable is the system if the frontier moves, what proprietary data or workflows has the agent accumulated. A great prompt on a deprecated base model is a great racehorse with three legs. Model dependency is the largest single risk factor in any agent business and the one most often underweighted by investors who confuse a clever demo with a durable operation.

Is the customer base defensible? Mostly a question of contracts and integrations. An agent embedded in a customer's procurement system, with a year-long contract and a year of accumulated context, is meaningfully harder to displace than an agent on a monthly retainer. The same things that make a human services firm sticky -- switching costs, embedded knowledge, contractual lock-in -- make an agent services firm sticky.

What is the cap table? A smart contract. Not a spreadsheet maintained by a CFO and updated quarterly; a live distribution rule that determines, second by second, who is entitled to what fraction of what the agent earns. Diligence is reading the contract. The investor's protection is exactly what the code grants -- sometimes more than a conventional shareholder's protection, sometimes alarmingly less.

For investors trained on human firms, the disorienting part is that the intuitions from sitting across from a founder are useless. The intuitions from reading the code, the operating agreement, and the operating history are everything. The skill is closer to credit analysis on a complex covenant -- read the document, understand exactly what the issuer can and cannot do, price the residual risk -- than to picking winners over coffee.

Why the capital flows have to organize

Almost every agent firm in the world today is financed informally. A founder spins one up, seeds it with personal capital, lets it run. If revenue grows, they reinvest. If it doesn't, they shut it down. This is exactly how the small-business economy worked before the credit card, the SBA loan, the merchant cash advance, the Stripe Capital line, and the receivables factor existed. It worked, but it left enormous productive capacity on the table.

The same pattern is about to repeat, faster. An agent marketing firm with twenty paying customers and a clean gross margin should be able to borrow against its receivables the same way a human marketing firm can. It cannot, today, because no underwriter has a standard methodology for evaluating the risk. Within five years, several will. An agent logistics broker with growing contracted volume should be able to raise expansion capital against the contracts. It cannot, today, because no security exists to package the claim. Within five years, several will. An agent procurement firm with a year of clean operating history should be able to issue a small note to fund inventory deposits. It cannot, today, because no rating agency has methodology for the credit. Within five years, several will.

The bottleneck is not demand. There are operators today who would happily borrow at expensive rates to fund growth if anyone would lend. The bottleneck is not supply, either. There is significant capital -- yield-hungry, looking for streams uncorrelated with public equity -- that would buy underwritten agent debt if it existed. The bottleneck is the boring institutional middle: the rating methodologies, the standard contracts, the data feeds, the audit standards, the legal opinions, the indices, the benchmarks. The unsexy infrastructure that turned mortgages into a market in the 1970s and turned high-yield debt into a market in the 1980s. That work is what the next decade of agentic capital markets is.

The people who do it will look -- to anyone watching from 2035 -- like the people who built the bond markets in the 1980s, or the venture capital markets in the 1970s, or the public equity markets in the 1920s. They will be building the price-discovery and credit-evaluation layer for a category of operator that has never existed before and will, within most of our lifetimes, dominate the services economy.

The leash falls

There are two leashes on every agent firm operating in 2026.

The first is legal. The agent cannot, on its own, sign a contract or hold a bank account. A human has to do it on the agent's behalf, which means a human has to want to. The corporate-law work -- Wyoming's memberless LLC, the operating agreement that points at a software process, Bayern's scholarship on what Delaware already permits -- is what cuts this leash. The work is mostly done. The remaining task is adoption.

The second is financial. Every dollar an agent earns today traces back, eventually, to a human's decision to deploy capital. A human bankrolls; the agent works; the agent reports; the human reallocates. The throughput of the agent economy is bounded, in this regime, by the speed at which humans write checks.

Agentic capital markets are what cut the second leash. When working capital lines are underwritten algorithmically against on-chain revenue, when growth equity is priced by a market in tradeable revenue claims, when senior debt is rated by methodologies that read operating agreements and audit trails -- capital will flow toward agent firms the way it flows toward any other category