Agents are not killing software.
They are killing the seat.

On September 8 the S&P 500¹⁹, the Nasdaq, and the Dow all closed lower. Application software led the way down. Reuters put the software and services index off about 1.4%⁰⁸ that day and listed AI worries next to Middle East risk and rates. So this was not a clean signal. It was a messy one.

The bigger number sits behind it. Software stocks shed roughly $2 trillion⁰² since their 2025 peak, according to an August 25 WebProNews report on what traders named the SaaSpocalypse. The iShares Expanded Tech-Software Sector ETF fell nearly 30%²² from its 2025 peak. For the first time, software forward P/E multiples as a group traded below the S&P 500¹⁹'s, per SaaStr analysis.

The market story is simple. Agentic models like GPT-6 Astra do multi-step work inside apps. If agents can do the work, software is cheap to build, so software is worth less. I think that story is half right, and the wrong half is the expensive one.

Here is my damaging admission up front. I cannot tell you how much of September 8 was agents and how much was rates or Iran. Nobody can. What I can show you is which part of the SaaS model actually breaks, why the login is the weak point, and what to change on your pricing page before a renewal forces you to.

The Three-Meter Contract

Per-seat pricing was never about software. It was a proxy for headcount. A March 2026 Juggerinsight analysis put it bluntly: every seat sold mapped to a human logging in, and that predictable relationship is what earned SaaS its premium multiples.

Agents break the proxy. The same analysis estimates one autonomous agent can carry the cognitive workload of two to three full-time equivalents. It drafts, schedules, updates the CRM, and never needs a license per user. The work still happens. The login does not.

So what replaces the seat? Not one thing. A July 31, 2026¹² review of eight large software platforms found that seats persist, but contracts are drifting toward a three-part structure. I call it the Three-Meter Contract.

Meter one is the platform commitment. That is your floor, the fee for being connected at all. Meter two is human licenses, priced for the people who still need identity, history, and a place to log in. Meter three is a shared agent allowance, a pool of tasks, actions, or credits that the agents burn down no matter who is at the keyboard.

The framework earns its keep by sorting products, not describing them. Collaboration tools like chat and documents show low to medium seat pressure, because every employee still needs an identity. Developer tools keep per-developer licenses and bolt on AI consumption. Back-office processing, meaning invoices, tickets, and documents, is where seats are getting crushed and per-workflow pricing takes over.

One more number to keep in your head. Simon-Kucher reported on March 17, 2026 that more than 75% of AI providers are unsure how to price their agentic products. That is your competition. Three quarters of them do not know what their third meter is yet.

Watch a 50-Seat Bill Fall Apart

Let me show you exactly how this hits a real bill. These numbers are illustrative, not from any filing. Picture a customer support tool at $100 per seat per month with 50⁰² agents on the floor. That is $5,000 a month, stupid easy to forecast, and the reason your investors loved you.

Now the customer deploys an agent that handles first-touch tickets. If that agent carries the load of two to three people, the customer does not need 50⁰² seats. Say they drop to 20. Your bill falls from $5,000 to $2,000, a 60% cut, while the volume of work running through your product went up.

That is the hard way. You keep the seat model, the customer keeps cutting seats, and you get repriced from the outside. The easy way is to meter the work before the customer does the math for you.

The model layer is already showing you the lazy way to do this. On the same September 8, OpenAI said ChatGPT Images generation latency was cut by up to 50%⁰² versus Images 2.0 and shipped two API tiers, GPT-Image-2.5 Flare and Flow. Read that closely. Unit cost went down and price segmentation showed up on the same day. That is what a healthy third meter looks like: as inference gets cheaper, you tier the work. You do not give it away.

Meta's move that day tells the other half. Its new agent reaches across applications to send email, list a car for sale, and book travel from one interface. Three standalone workflows collapse into one prompt. The human never opens the travel app, so the travel app never gets to count a seat. If the only way you bill is the login, you just got billed out of the flow.

Now the part most pricing hot takes skip. Do not overcorrect into pure outcome pricing. If your outcome depends on the customer's messy data, you eat the loss.

There is a revenue side too. Once agents run on their own logic, usage can spike without a human touching anything, which means surprise bills and quarter-to-quarter revenue swings that public markets punish. CFOs still budget per person and per cost center. The winning third meter is usually a pooled allowance with a hard cap and a predictable overage, not a naked per-token charge.

So cut the noise. Ignore the debate about whether seats are dead. Keep seats for humans, keep the platform fee as your floor, and put a capped, pooled agent meter on the one workflow where your customers are already cutting heads. That is the whole play.

What breaks when the login disappears

2031. Pull back five years. Two things are true at once. Capital is fleeing the application layer, and capital is flooding the model layer. Mistral just raised €3 billion at a €21 billion post-money valuation while application software got marked down. Investors are not saying software is over. They are saying the surplus moved.

One contrast worth memorizing. Seats bill for presence. Meters bill for work. Presence is what agents remove. Work is what agents multiply. A pricing model tied to presence has capped upside and uncapped downside, which is exactly the asymmetric bet you do not want to hold.

There is a real bear case, and I want to hold it honestly. An EQT analysis from October 23, 2025¹⁵ argues agents from OpenAI, Meta, and Alphabet might eventually complete tasks directly themselves and skip the human-facing app entirely. Deloitte's July 22, 2026¹⁶ piece goes further: interfaces shrink into control centers for agents, and orchestration layers could commoditize the applications underneath. In that world you fix pricing and still lose, because you became a backend.

I do not know whether that happens broadly or only in narrow back-office lanes. The data is mixed. My read is that the commoditization risk is real for undifferentiated workflow tools and overstated for anything that owns a system of record.

Think of Costco's $1.50¹⁸ hot dog. It is not the profit center. It is the anchor that keeps you walking in. Your platform commitment is the hot dog: cheap, stable, and the reason the agents route through you rather than around you. The margin lives in the third meter.

The 2031 winners will have practiced beginner's mind about their own contracts. They will have accepted that the seat was never permanent before the market forced it on them, then counterpositioned against incumbents who cannot lower per-seat prices without blowing up guidance. Whoever has the cleanest agent meter by 2028 will own the renewal conversation in 2031.

Put an Agent Meter on One Plan

You do not need a pricing consultant or a CS degree for this. You need one workflow and a spreadsheet. Get your reps in.

First, pick the single workflow where your customers are already replacing people with agents. Invoices, tickets, lead research, and drafting are the usual suspects. If you sell into go-to-market teams, Nex is a Claude-powered coworking layer built for repetitive work like lead research and sequencing. Scope it to one high-volume workflow and watch how many human touches disappear. That count is your new unit.

Second, define the work unit in words a CFO can budget. "Per resolved ticket" or "per 1,000 enriched leads" beats "per token" every time. A unit is just the smallest thing your customer would happily pay for again. If you introduce a meter nobody can predict, you have built a surprise-bill machine.

Third, build the three-part offer. Platform fee as the floor, human seats for whoever still logs in, and a pooled agent allowance with a hard cap. MagiCrew pools agents into one shared workspace, but it encourages you to define roles before you deploy agents widely. Steal that discipline. Define the roles your agents play for the customer before you decide how to charge for them.

Fourth, test the meter on five customers, not fifty. Tell them plainly you are testing. Ask them to poke holes until you are confident the cap and overage make sense. Revolte pitches itself as a pipeline from intent to production that plans and runs checks. Treat your pricing test the same way: state the intent, then audit every check the meter runs against a real invoice.

Fifth, expect it to break. Your first unit will be wrong. Your first cap will be too low or too high. That is normal, and it is a lot cheaper than learning it from a lost renewal. Ship the meter, learn in public, and fix it in a matter of weeks rather than waiting for the market to fix it for you on a day like September 8.