brianletort.ai
All issues

The Application Layer

Issue 16 · Week 33 of 2026.

/Weekly read/~7 min read/Public sources onlyDownload brief

The Big Read

Two opposite answers shipped in three days to the same question: does an agent come in through the front door as a user, or get assigned work inside the system of record?

The thesis this issue defends

On August 11 SpaceXAI launched Grok Bot in early beta — agents that run on a cloud Linux VM, log into real applications through the browser, and do work in the user interface. No MCP server, no API integration, no vendor partnership. On August 13 Adobe pushed AI Collaborators and Task Agents into the Workfront production path, where an agent is a thing you assign a task to inside the work system, inheriting that system's permissions, audit trail and approval routing. Both are agents doing knowledge work. They are architecturally opposite, and the difference is entirely about who owns the permission.

That distinction is the procurement decision of the quarter, and it is not being framed that way anywhere. The front-door model is enormously attractive because it works with every application you already own, including the ones with no API and no agent roadmap. It is also, structurally, an identity problem: SpaceXAI's own documentation notes that a user's bots share the computer and its sessions, which means credential pooling across tasks is a default rather than an exception. A bot name is not a security boundary. If you are piloting this against a CRM or a finance system, the questions to answer before the pilot are which identity the agent authenticates as, whether its credentials are isolated per task, and what your SIEM sees when it acts — because from the application's perspective, it sees a user.

The assigned-worker model gives up reach and gets governance. An agent assigned a Workfront task cannot do anything the assignment does not permit, and the work shows up where the rest of the work already is. Adobe's version is telling in the detail: the feature carries no additional Workfront fee, but third-party agent usage may still bill separately, and the task agents connect out to Claude, Copilot Studio and Writer. So the incumbent kept the permission model and the workflow surface, and let the model layer be someone else's cost line. That is the incumbent SaaS playbook working exactly as the thesis predicts — the system of record does not need to win the model, it needs to remain the place where work is assigned.

The week's most useful number came from OpenAI and cuts across both models. Its enterprise telemetry shows the top decile of customers by output tokens per active user running at 8.3x the intensity of typical firms in June, against 2.6x in January. Among weekly active users, plugin use is 21% versus 9% and skills use is 19% versus 3%. Read that carefully, because it is vendor telemetry on a vendor's own definition and it still says something hard to dismiss: the gap between organizations is widening roughly three times faster than any gap in model access. Every firm in that comparison can buy the same models at the same prices. What separates them is whether workflows have been packaged into reusable skills and connected to real systems. CIOs still reporting AI progress as percentage of seats licensed are measuring the input that stopped being the constraint.

One structural change worth naming plainly: SpaceX closed its acquisition of Cursor on August 14, and with it the clearest independent coding-agent pure-play is gone. Coding agents are now overwhelmingly owned by companies that also own frontier models and, in this case, compute. For CIOs the question is no longer which coding agent is best but how much concentration you are willing to hold across model, tool and infrastructure with one counterparty. For founders in adjacent categories, the exit comparison just got set by a buyer with a strategic rather than financial rationale, which raises the price and narrows the buyer list at the same time.

Vertical AI itself was quiet. No funding round, named customer win, or M&A landed in-window from Harvey, Abridge, Sierra, Hebbia, Glean, EvenUp or Ambience, and this issue does not manufacture one. That silence is worth a note rather than a paragraph: in a week when the horizontal platforms shipped two new agent form factors, a pricing halving, and an acquisition close, the vertical cohort produced nothing public. Two or three more weeks of that pattern would be a real signal about where the value is accruing.

Vertical movements

Vertical packages shipped.

5 vertical packages shipped this period.

Industry- or function-specific application packages that landed this period — from frontier labs, open-weights labs, and the insurgent vertical-AI startup cohort. Each entry names the vertical it serves and the decision implication for buyers of the displaced tier.

  • /SpaceXAI/Operations/Frontier lab/Seat

    Grok Bot

    Early-beta computer-use agents that run on a cloud VM and log into existing applications through the browser, with no MCP server or API integration required

    CIOs should treat this as a new procurement category rather than a new feature, and require identity, credential isolation, audit and approval policy before any pilot touching CRM, finance or HR. The vendor's own documentation notes a user's bots share the computer and its sessions, so credential pooling is the default and a bot name is not a security boundary. For vertical-AI founders in Sales and Support, the 'only we can finish the last mile in the UI' differentiator just weakened considerably.

    SpaceXAI, VentureBeat, The Verge

  • /OpenAI/Security/Frontier lab/Usage

    GPT-5.6-Cyber and the Daybreak Blue / Red tiers

    A purpose-trained defensive security model distributed only through a vetted access tier at $12.50 / $75 per 1M tokens, with hardware security keys required from September 1

    Security leaders planning to use frontier models for vulnerability research should start vetting and key provisioning now rather than at point of need, because the gate here is organizational rather than technical and the September 1 date is firm. The more consequential precedent is commercial: this is the first vertical package where the access-control design, not the capability, is the product. Buyers of security tooling should expect capability tiering by customer vetting to spread to other dual-use categories.

    AI/TLDR, DEV Community, The Neuron

  • /Google/Engineering/Frontier lab/Usage

    Gemini 3.7 Flash

    Agent-oriented workhorse model at $0.75 / $3.75 per 1M tokens, half the prior Flash rate, shipped GA across the Gemini API, Enterprise Agent Platform and Spark, and into GitHub Copilot the same day

    Engineering leaders should re-run agent TCO models this week and then immediately note the expiry: Google's own pricing page states the rate is introductory through December 31 and becomes $1.50 / $7.50 on January 1. Teams on GitHub Copilot get the model without a new vendor contract but Business and Enterprise tenants must enable it by policy, so update model allowlists and cost alerts before defaults drift. Vertical founders whose margin depends on wrapping cheap Flash-class inference should assume that arbitrage has a published end date.

    Google, Google Cloud pricing, GitHub Changelog, InfoWorld

  • /Adobe/Marketing/Incumbent SaaS/Hybrid

    Workfront AI Collaborators and Task Agents

    Agents become assignable workers inside the work management system, connecting out to Claude, Copilot Studio and Writer, with MCP client actions from ChatGPT, Claude and Copilot

    Marketing and operations leaders should prefer agents assigned inside the system of record over side-chat pilots, because the assignment inherits the permission model, the audit trail and the approval routing you already run. Note the commercial structure: the Workfront capability carries no additional fee while third-party agent usage may bill separately, so the incumbent keeps the workflow surface and pushes the model cost to another line. Vertical-AI vendors should plan to integrate as assignable workers rather than build a parallel interface.

    Adobe Experience League, Computerworld

  • /GitHub and coalition partners/Engineering/Incumbent SaaS/

    Agent Plugins 1.0

    A portable plugin package bundling skills and MCP servers, now GA across VS Code, Copilot CLI and the Copilot app, with a spec backed by AWS, Anysphere, Microsoft, OpenAI, Vercel and Google

    Platform teams should standardize on one plugin package per capability and ban per-IDE forks now, while the format is young enough that consolidation is cheap. Govern it as software supply chain rather than as configuration: enterprise controls run through existing managed settings and MCP allowlists, which means the security review belongs to whoever already owns endpoint policy. The multi-vendor backing is the notable part — this is the first agent extensibility format with enough signatories to plausibly outlive any one vendor's tooling.

    GitHub Changelog

Incumbent responses

How the SaaS estate is answering.

4 incumbent SaaS responses worth tracking.

Established SaaS vendors reacting to the agentic shift — product launches, repositioning, earnings color, partnerships, and restructuring. The system-of-record incumbents defending their turf against systems of action.

  • /Microsoft/Copilot app consolidation and the path to Autopilot

    Consumer and work Copilot apps begin merging on mobile and web with desktop following mid-September, while Group Chat, Podcasts and Deep Research retire from August 18 and the promised agent 'super app' remains unshipped and unpriced

    Enterprise buyers should map SKUs before Q4 budgeting and negotiate agent meters separately from the $30 per user Copilot seat, because seat Copilot, Cowork credits and a future Autopilot tier are three different commercial instruments being presented as one product family. Retiring shipped consumer features while promising an agent super app in the same quarter is a signal about where engineering attention has moved, and buyers renewing this quarter should ask for pricing protection on capabilities that do not yet have a price.

    Microsoft Support, TechCrunch, GeekWire, Computerworld

  • /Adobe/Workfront AI Collaborators and Task Agents

    Agents become first-class assignable workers inside the work management system on the fast release track, with full production availability for all customers on October 15

    This is the clearest expression yet of the incumbent defensive play the thesis predicts: do not try to win the model, remain the surface where work is assigned and permissioned. For SaaS investors it argues that systems of record with real workflow state are more defensible against agentic disruption than systems of engagement. For CIOs it means the agent governance conversation can ride on the existing application's permission model rather than requiring a new one.

    Adobe Experience League, Computerworld

  • /Oracle/OCI day-zero Nemotron 3.5 Lightning and Responses API background mode

    Day-zero open-model availability plus a background execution mode for long-running agentic tasks, both inside the existing OCI identity boundary

    Enterprises already on OCI can route agent workloads to a cheaper open model without leaving their IAM boundary, which removes one of the stronger arguments for a third-party agent infrastructure vendor. SaaS investors should assume incumbent clouds keep absorbing the agent-runtime layer that independent vendors were charging for; the margin in 'we run your agents' is compressing toward zero wherever the customer already has a cloud contract.

    Oracle AI blog

  • /GitHub and Microsoft/Copilot for JetBrains persistent memory with enterprise managed settings

    Cross-session agent memory shipped alongside managed settings that let administrators lock MCP servers, plugin marketplaces and permissions

    Engineering leaders should enable memory only with a retention and redaction policy in place, and should lock MCP servers and plugin sources through managed settings before agents gain cross-session recall. The pairing is the notable part: the vendor shipped the governance control in the same release as the capability that requires it, which is a pattern worth demanding from every agent vendor rather than treating as generous.

    GitHub Changelog

Startup signals

The insurgent vertical cohort.

3 startup signals this period.

Vertical-AI startups raising capital, winning named customers, or shipping general-availability product. The cohort that sits between frontier labs moving down the stack and SaaS incumbents defending their record-of-truth.

  • /Cursor (Anysphere)/Engineering/Widely reported at ~$60B; not stated in the company's own announcement

    SpaceX closed its acquisition of Cursor; the product continues with its compute and model roadmap now aligned to SpaceXAI

    The clearest independent coding-agent pure-play is gone, and CIOs should re-run concentration risk rather than tool evaluation: model, agent and now compute sit with one counterparty. For investors the category just lost its reference comparable for an independent exit, and the price was set by a strategic buyer with a compute thesis, which raises the number and shortens the buyer list simultaneously. Treat the deal size as reported rather than confirmed until it appears in a filing.

    Cursor, The Verge, Business Insider

  • /Cursor (Anysphere)/Engineering

    Earned AIUC-1 certification following a Schellman audit with adversarial testing, with quarterly retesting required

    Procurement should start asking coding-agent vendors for agent-behavior attestations rather than accepting SOC 2 as sufficient, because SOC 2 says nothing about what an autonomous agent will do with the access you granted it. Adding AIUC-1 or an equivalent to the standard security questionnaire is a low-cost change that materially sharpens vendor comparison this quarter, and early certification is currently a differentiator rather than table stakes.

    Cursor, AIUC-1

  • /Together AI/Engineering/$240M

    Signed a multi-year agreement with IBM for an NVIDIA HGX B300 inference cluster on IBM Cloud, roughly 2,000 Blackwell chips available Q1 2027

    Confirms capital is rotating into inference offtake and open-model serving rather than only frontier training, and that mid-tier clouds are now writing nine-figure GPU commitments to secure it. For founders building on open weights, serving capacity at this scale becoming contractible through a mainstream enterprise cloud lowers the infrastructure barrier — and raises the question of what remains defensible once serving is a commodity procurement.

    IBM Newsroom, Reuters

Pricing shifts

Seat to outcome, one move at a time.

3 pricing-model shifts announced.

Public pricing-model shifts inside the window. The 'data owns the application' thesis predicts a structural move from seat-based to outcome-based pricing across SaaS; tracking the rate of change is itself a market signal.

  • /Google

    Usage-basedUsage-based

    Gemini 3.7 Flash launched at $0.75 input / $3.75 output per 1M tokens, half the prior Flash generation, with Google's own pricing page stating the rate is introductory through December 31, 2026 and that $1.50 / $7.50 applies from January 1, 2027. The shift being tracked is not the level but the structure: a published expiry date on a promotional inference rate, disclosed at launch rather than discovered on renewal.

    Google Cloud pricing, Google blog

  • /Microsoft

    Seat-basedHybrid

    The Copilot app consolidation makes the commercial structure visible: a $30 per user Copilot seat, separately metered Cowork credits, and a promised Autopilot agent tier with no announced price. Buyers renewing this quarter are being asked to commit to a seat while the agent capability moves to a meter that has not been published.

    Microsoft Support, Computerworld, GeekWire

  • /Adobe

    Seat-basedHybrid

    Workfront AI Collaborators and Task Agents carry no additional Workfront fee, but third-party agent usage through connected Claude, Copilot Studio or Writer endpoints may bill separately. The incumbent holds the seat price flat and lets the variable cost land on the model vendor's meter — a structure that makes the agent feature look free while moving the marginal cost off the SaaS invoice.

    Adobe Experience League, Computerworld

Vertical scorecard

Who leads each vertical.

10 verticals · leaders as of Aug 15, 2026.

A snapshot of leader-vs-challenger by vertical. Useful for procurement shortlists when matching workload to vendor cohort. Rows refresh weekly as leadership shifts.

  • Legal

    Leader: Harvey

    Challenger: Legora

    Unchanged for a second consecutive week with no qualifying in-window movement. The category still has no independent matter-level ROI study, which remains the gap that would settle it rather than another product announcement.

  • Security

    Leader: OpenAI (GPT-5.6-Cyber via Daybreak Red)

    Challenger: Salesforce (Agentforce 360 at IL5)

    Position changes on capability-plus-gating: a purpose-trained model for vulnerability discovery, distributed under vetting with hardware keys from September 1, is a different category of offering than an accredited agentic portfolio. Salesforce keeps the regulated-deployability lead; OpenAI takes the capability lead.

  • Engineering

    Leader: Anthropic (Claude platform)

    Challenger: SpaceXAI (Cursor plus Grok 4.6)

    The challenger position changes hands on the Cursor acquisition close rather than on capability. Anthropic holds the runtime lead, but the combination of a leading agent surface, a frontier model at $2 / $6 and captive compute is the most vertically integrated engineering stack any vendor now fields.

  • Finance

    Leader: Microsoft Dynamics

    Challenger: Oracle Fusion Agent Studio

    Unchanged on product. Oracle's day-zero open-model availability and background execution mode strengthen the challenger's infrastructure story without shipping a finance-specific capability, so the position holds.

  • Support

    Leader: Salesforce (Agentforce)

    Challenger: Sierra (Horizon)

    Unchanged, but Grok Bot is the first credible threat to both from outside the category: a computer-use agent that operates the existing support console needs no vendor to build an integration and no customer to migrate.

  • Commerce

    Leader: Salesforce (Agentforce Commerce)

    Challenger: OpenAI (ChatGPT checkout)

    Unchanged; transaction ownership remains decisive and no in-window commerce product event moved the position for a third consecutive week.

  • Operations

    Leader: Microsoft (Copilot Studio)

    Challenger: SpaceXAI (Grok Bot)

    The challenger changes on form factor. Copilot Studio requires connectors and a build step; Grok Bot claims to operate any application a human can, which is a fundamentally wider reach at a fundamentally weaker governance posture. Reach wins pilots, governance wins renewals.

  • Research

    Leader: OpenAI (Academic Researchers)

    Challenger: Google (Gemini Enterprise Agent Platform)

    Unchanged on product. Google's rapid Flash cadence strengthens the platform argument while its unannounced Pro timeline weakens the frontier-research argument, leaving the position where it was.

  • Marketing

    Leader: Adobe

    Challenger: Salesforce

    Adobe extends its lead on the strongest structural move in the category this week: agents assignable inside the work system of record, inheriting its permissions, at no additional platform fee. The model cost lands on a third-party meter, which keeps the seat price intact.

  • Other

    Leader: Snowflake (Cortex AI Gateway)

    Challenger: Provider-native MCP surfaces

    Unchanged, and the gateway case strengthened again this week: with DeepSeek swapping a served build without notice, the ability to pin and disclose which build and endpoint served a request is becoming a governance requirement rather than a convenience.

Architecture watch

Patterns to track.

4 cross-vendor patterns reshaping the application layer.

Patterns that crossed multiple vendors this period. One pattern, several exemplars, what it changes for procurement, cost, or vendor-risk posture.

  • Front-door agents versus assigned agents — the permission fork

    Grok BotAdobe Workfront Task AgentsCopilot Studio

    Two architecturally opposite agent form factors shipped within three days. The front-door model runs on a VM and operates applications through the browser as a user, which gives it reach into every system you own including those with no API — and makes it an identity problem, since the vendor's documentation notes bots share a computer and its sessions. The assigned model makes the agent a worker inside the system of record, which caps its reach at what the assignment permits and inherits the existing audit trail. Buyers should decide which model they are adopting deliberately, per system, and write the identity answer down before the pilot rather than after it.

    SpaceXAI, Adobe Experience League, Computerworld

  • Depth, not seats, is the adoption metric that separates organizations

    OpenAI Enterprise SignalsAgent Plugins 1.0Copilot memory

    OpenAI's telemetry puts its top-decile customers at 8.3x the output tokens per active user of typical firms in June against 2.6x in January, with plugin use at 21% versus 9% and skills at 19% versus 3%. It is vendor data on a vendor's own definition, but the direction is corroborated by where the tooling went this week: portable plugin packages, persistent memory, managed settings. Every firm in that comparison buys the same models at the same price, so the widening gap is organizational reuse rather than access. CIOs reporting progress as percentage of seats licensed are tracking an input that stopped binding.

    OpenAI Enterprise Signals, GitHub Changelog

  • Incumbents keep the permission surface and push model cost to someone else's meter

    Adobe WorkfrontOracle OCIMicrosoft Copilot tiers

    Adobe shipped assignable agents at no additional Workfront fee while third-party agent usage bills separately. Oracle added day-zero open-model serving inside its own IAM boundary. Microsoft is separating a seat price from Cowork credits from an unpriced agent tier. The common structure is that the incumbent retains the thing that is hard to replicate — workflow state, identity, the place work is assigned — and lets the model layer be a variable cost carried elsewhere. For SaaS investors this argues systems of record with real workflow state are more defensible than the agentic-disruption narrative assumes.

    Adobe Experience League, Oracle AI blog, Microsoft Support

  • Cheap inference is being sold with a published expiry date

    Gemini 3.7 FlashDeepSeek V4 ProGrok 4.6 promotional period

    Google halved Flash pricing and stated at launch that the rate expires December 31 and doubles January 1. DeepSeek's flat rate gives way to peak and off-peak tiers on August 16. SpaceXAI shipped Grok 4.6 with a one-week 2x usage promotion. Vendors have moved from discovering price changes on an invoice to publishing them in advance, which is better behavior and also a clear statement that current agent unit economics are promotional. Any vertical-AI business case whose payback crosses January needs the step-up modelled, and any pricing built on token arbitrage needs a second source of margin.

    Google Cloud pricing, OpenLLMStack, SpaceXAI

Watchlist

On the radar next.

6 catalysts to watch, starting Aug 16.

Forward catalysts in the next 7–30 days that would change the read materially — earnings prints, conferences, expected product launches, regulatory decisions, and competitive responses.

  • Aug 16

    DeepSeek moves to peak and off-peak API pricing

    The first major inference provider to price by time of day. If application vendors start scheduling batch agent work into off-peak windows, time-of-day arbitrage becomes a real architecture discipline and a line item in vertical-AI gross margin rather than an infrastructure curiosity.

  • Aug 18

    Microsoft retires consumer Copilot features as the apps merge

    Group Chat, Podcasts and Deep Research retire, with desktop consolidation following mid-September. Enterprises should finish SKU mapping before this lands, because the retirement list is the clearest available signal of which capabilities Microsoft will and will not carry into the Autopilot tier.

  • Sep 15 - Oct 15

    Adobe Workfront Task Agents reach full production for all customers

    October 15 is the general availability date. Watch whether the third-party agent billing is itemized transparently, because that determines whether the assigned-worker model is genuinely cheaper for buyers or simply moves the same cost to a meter they read less carefully.

  • September

    Microsoft Autopilot pricing disclosure

    The agent super app was promised this quarter and remains unpriced. Whether it lands as a seat uplift, a credit meter, or an outcome-based SKU is the single most consequential pricing decision in enterprise software this year, given the installed base it would apply to.

  • Aug 16 - Sep 30

    Vertical-AI cohort funding and customer-win silence

    No round, named win or M&A landed in-window from the major vertical-AI names. Two or three more weeks of the same while horizontal platforms ship agent form factors would be a genuine signal that value is accruing to the platform layer rather than the vertical applications built on it.

  • Q4 2026

    Agent-behavior certification becomes a procurement requirement

    Cursor's AIUC-1 certification with quarterly retesting is currently a differentiator. Watch whether a second major agent vendor certifies before year end — that is the point at which security questionnaires start requiring it and the absence of certification becomes disqualifying rather than neutral.

Edits this issue

  • Vertical-AI startup silence is reported as a finding rather than filled. No qualifying in-window funding, customer win or M&A was found for the major vertical-AI names, and the issue says so in the big read instead of promoting near-window items into the section.
  • The scorecard's security, engineering and operations rows change on structural moves — capability gating, an acquisition close, and a new agent form factor respectively — rather than on benchmark movement.
  • Pricing shifts now record structural changes as well as level changes: a published expiry date on a promotional rate is tracked as a pricing-model event even when the model category is unchanged.

About The Application Layer

A weekly read on the layer above the model — vertical packages from frontier labs, incumbent SaaS counter-attacks, vertical-AI startup signals, and pricing-model shifts. Sibling to The AI Stack Weekly (the cross-stack flywheel) and The Model Pulse (the model layer).

Authorship and sources

Compiled from public vendor announcements, SEC filings, earnings releases, conference coverage, and reputable trade press. Written by Brian Letort. Independent analysis. Not investment guidance.

Operate. Publish. Teach.