LAYER 06 / 06
Layer 06 — Applications & Agents

Does off-the-shelf survive?

This is the layer that has to ultimately justify everything spent below it. The macro question for the entire sector eventually resolves here: is there enough real, paid value at the top to justify the capex at the bottom?

The 101

What this layer is

This is what end customers actually pay for: the copilots, vertical suites, and autonomous agents doing real work. Every other layer of the stack — the land, the silicon, the data centers, the models, the harnesses — exists in service of the revenue that lands here.

The central problem of the layer is the thin wrapper. A product survives only by adding something the model alone cannot provide: proprietary data, deep workflow integration, distribution, trust and compliance, or a system of record. Value accrues to whoever owns the customer relationship, the proprietary data, and the workflow — not whoever has the cleverest prompt.

The defining shift

Pricing moves from seats to outcomes

In three years the dominant billing unit moved from the seat to the outcome — thirteen dated public events trace the shift.

Per-seat → per-outcome, 2023–2026
13 dated pricing-model shifts · public announcements
2023 2024 2025 2026 2023-03 · INTERCOM LAUNCHES FIN AT $0.99 PER RESOLUTION 2024-08 · ZENDESK: OUTCOME-BASED PRICING, PAY PER RESOLUTION 2024-09 · SALESFORCE AGENTFORCE: $2 PER CONVERSATION 2024-12 · COPILOT STUDIO: $0.01 PER MESSAGE, PAY-AS-YOU-GO 2024-12 · DEVIN SHIPS METERED IN ACUS FROM DAY ONE 2025-04 · DEVIN 2.0: $20 ENTRY, PAY-AS-YOU-GO ACUS 2025-05 · SALESFORCE FLEX CREDITS: $0.10 PER ACTION 2025-06 · CURSOR PRO GOES USAGE-BASED, API-PEGGED 2025-06 · GITHUB COPILOT METERS PREMIUM REQUESTS 2025-07 · CLAUDE CODE: WEEKLY CAPS, EXTRA USAGE AT API RATES 2026-02 · HIGHRADIUS DROPS PER-SEAT ENTIRELY (REPORTED) 2026-04 · SERVICENOW: CONSUMPTION-BASED AI TIERS 2026-06 · SALESFORCE AGREES TO BUY FIN FOR $3.6B OUTCOME PRICING GETS ACQUIRED

Hand-curated from vendor pricing pages, press releases, and dated coverage (Stripe, Zendesk, Salesforce, Microsoft, TechCrunch, GitHub docs, and others); committed as data/apps-pricing-timeline.json. Cursor and Claude Code moved to usage-billed rather than true per-outcome pricing; the HighRadius entry is single-source and presented as reported.

The hardest evidence that this layer is being rebuilt around agents is how it charges: from the seat — a human logging in — toward the outcome, a resolved ticket, an action taken, a unit of agent work. The last event on the timeline is the loudest — the biggest per-seat vendor in software agreed to buy the per-outcome pioneer for $3.6B.

Follow the dollar

Where $1.00 of AI application spend lands

$0.14
Apps + Agents
$0.10
Harness
$0.18
Models
$0.30
Data Centers + Compute
$0.22
Silicon
$0.06
Land + Power + Shell

Apps and agents take $0.14 of the illustrative dollar — and this is the slice that has to justify all the others. The gap is stark: Menlo Ventures estimates roughly $19B of enterprise AI application revenue in 2025, against an estimated $700–900B of 2026 hyperscaler capex — roughly 35–45x. Both figures are public estimates, and the application number excludes consumer spend, so the true gap is somewhat narrower.

Illustrative split — assumptions in Methodology.

Deep dive

From selling seats to selling work

Off-the-shelf melts into custom

Chamath Palihapitiya argues that the differentiation between off-the-shelf and custom software melts away in this era: with the right harness, every company can imbue its own alpha into the software that runs it, and off-the-shelf is largely replaced by custom software. In his framing, build-once-sell-repeatedly becomes a laggard go-to-market motion for a SaaS world this stack no longer needs.

"Custom by design, alpha embedded, proprietary by nature." — Chamath Palihapitiya

That remains his claim, made from a single vantage point. The pricing timeline above is the hard evidence; the quote is the color. What the timeline shows is vendors themselves conceding that the seat was never the thing being sold.

Pricing follows the labor

As agents do actual labor, the billing unit follows the work: per resolution, per action, per unit of agent effort. This expands the market — software priced against labor budgets rather than software budgets — and opens a new battleground, because outcome pricing forces vendors to eat the risk of their own product failing. Sierra has charged this way since inception, at a reported $1.50 or so per resolved interaction. The strongest single datapoint is the exit: Salesforce, the biggest per-seat vendor in software, agreed to acquire Fin — the $0.99-per-resolution pioneer — for $3.6B in June 2026. The things to watch now: paid adoption versus pilots, agents crossing from demo to production, and the net revenue retention of AI-native software.

The agentic-commerce endgame

Play the tape forward on agents doing consumer work. You want a pizza; your app tells your agent through a model API; on the other side, another agent finds a restaurant, builds the order, and pays. Every step of that chain is invisible to the browser and the storefront. Whoever holds the credential and the payment holds the customer relationship — the app-layer prize migrates to whoever intermediates the agents. That is Stripe's bet: humans interact through agents, tool-calling gets absorbed into APIs, and the neutral intermediary routes everything, payments included.

The agent internet diverges from the human internet

Agents browse nothing; they call. The Orthogonal team argues on the Stateful podcast that the human and agent internets are diverging: applications must expose structured, per-call access — pay-per-call micropayments instead of $250-a-month subscriptions — and providers are already paying to be discoverable in front of agents, a dynamic they call GEO versus SEO. In their view most transactions eventually flow through agents, with guardrails and trust maturing first through small API-level purchases before larger ones. These are their claims, but they describe a distribution economy forming at this layer right now.

Whatever you build, you own

The business angle cuts the other way too. Software you imbue with your own alpha is yours entirely: you can run it, license it out, or sell the agent itself. The agent becomes the product — which means the layer's endgame is less a shelf of SaaS logos and more a market of owned, proprietary agents trading work with each other.

Consumer AI apps: privacy as the differentiator

The consumer end of this layer is consolidating around a few default assistants — which makes differentiated challengers instructive. Venice's consumer products (its private chat app, Characters, and Agentic Chat) compete here on a single promise the defaults can't easily copy: the conversation is encrypted end-to-end and never stored. It's a working example of this page's moat rule — distribution and trust beating raw capability — and a reminder that one company can span layers: Venice's home on this map is the Harness, where its privacy routing lives; its consumer apps are that membrane worn as a product.

Key players

Who's in this layer — and how they differ

HarveyVertical AI for law

Domain depth as the moat — legal workflow, trust, and elite-firm distribution that a generic model cannot supply.

SierraAgents as the product

Customer agents outcome-priced from inception — a reported $1.50 per resolved interaction, no seats anywhere in the model.

OpenEvidenceVertical AI for medicine

Clinical answers grounded in licensed medical literature, free to verified clinicians — the trust, sourcing, and physician distribution a generic model cannot supply.

VenicePrivate consumer AI · home layer: Harness

Consumer products — private chat, Characters, Agentic Chat — competing on a promise the default assistants can't easily copy: end-to-end encrypted, never stored. Cross-layer reach from its Harness home: the membrane worn as a product.

Vertical AI buildersCategory

Products that win on proprietary data and workflow depth in one domain — the moat the thin wrapper never had.

Incumbent app suitesCategory

Installed distribution and systems of record retrofitting agents into existing products — the per-seat world defending itself.

My read

Off-the-shelf software was always a compromise: we accepted average-fit tools because custom cost too much. That constraint is dissolving — with a real harness, a company embeds its own data, rules, and workflows into software that fits nobody else, at prices that used to buy a subscription. Durable winners here hold something a raw model can't replicate: distribution, proprietary data, deep workflow integration. And watch the pricing: the moment software charges per outcome instead of per seat, it stops being a tool and starts being labor. That repricing is the biggest market expansion I expect to see in my career, and it's already underway.

— Conner Murphy