In early 2026, a collective panic swept through software companies and repriced the entire category. The Dow Jones U.S. Software Index (DJUSSW index), which serves as a U.S. software sector benchmark, fell from a record close of $9040.04 on October 28^th, 2025, to $5773.08 on April 10, 2026 (a drawdown of roughly 36%).¹ By August 31^st, 2026, it had rebounded about 42.5% off that low.¹ What drove a swing that volatile, and have we traveled from a software apocalypse back to a software renaissance?

Our view: software as a whole was genuinely marked down by the "artificial intelligence (AI) eats software" narrative, but the more useful question is what is happening inside software, and, more fundamentally, who the user of software is becoming. Over the medium- to long-term, we believe agents will sit alongside humans as co-users of software.

The applications that both humans and agents can operate, call them agent environments, may be the winners of the new workflow. And the software that gives agents the ability to pay, communicate, remember, authenticate, and be authorized could become the agent stack's infrastructure layer, and therefore its bottleneck once agent deployment scales. We believe the KraneShares Public-Private AI & Technology ETF's (AGIX) latest rebalance reflects both convictions. Below, we walk through the reasoning.

Possession-Like Software As A Service (SaaS)

Set aside stocks and portfolios for a moment and look at AI agents from a purely futurist vantage point. Suppose artificial general intelligence (AGI), which is a hypothetical form of AI capable of learning, reasoning, and performing a broad range of intellectual tasks at a human-like level, already existed. What would it look like? What avatar would it wear, and how would it reach us?

Hollywood could have been answering this question for years. Jarvis in Iron Man. The voice in Her. A calm presence, a strong sense of being there, a phone number. The implication is that an agent does not need to be a new app or a new body. It needs presence and to hold all of our context. Its defining property, in other words, may be that it is possession-like; it inhabits the interfaces humans already use rather than asking for one of its own.

If that is right, the software we know best and use most often becomes the vessel for a possession-like intelligence. In this case, the agent never makes you leave your working environment, and never makes you re-explain who you are, what the company does, or who the customer is. Anthropic's (1.68% weight in AGIX as of 8/31/2026*) release of Claude in Slack is an expression of exactly this product taste. By "@-mentioning" Claude inside a channel, a user skips the tedium of pasting context and simply hands over the task. The agent treats the channel itself as its unit of work, continuously absorbing discussion, tracking task state, learning who is who, and, in some cases, acting asynchronously on its own initiative.

The trend appears to be showing up in results. On Salesforce's FY27 Q2 earnings call, management reported that Agentforce and Data ARR grew more than 200% to $3.9 billion, with Agentforce ARR up 200% to $1.5 billion (ARR, or annual recurring revenue, being the run-rate measure software companies use to describe subscription scale).² Before AI agents, Slack was a communication tool. Now it looks more like a canvas on which humans and agents collaborate.

Meanwhile, Salesforce (1.68% weight in AGIX as of 8/31/2026*), the previous era's system of record and the authoritative database of customers and deals, is abstracting that record into agent context through what it calls headless add-ons. Headless here means users do not need to work inside the full Salesforce CRM interface at all; they can reach Salesforce data, workflows, and agent capabilities through Slack, Teams, Coworker, Claude, or a company's own custom front end. That shift could extend Salesforce's reach to finance, operations, supply chain, product, and executive users who never needed to buy a CRM seat, potentially expanding the addressable user base inside existing accounts.

Twilio (1.64% weight in AGIX as of 8/31/2026*), the communications platform that lets developers send texts, place calls, and run chat through a few lines of code, is pursuing a parallel idea. Its agentic platform aims to free customer engagement from siloed channels and turn it into a conversational context that persists across all touchpoints. For consumers, the most natural agent will probably not ask them to download anything. As we argued above, it should attach itself to phone calls, SMS, WhatsApp, and web chat, the media people already live in, and become possession-like.

A daily communication with agents is demonstrated below:

Openclaw, the open-source (software whose underlying code is publicly available for users to inspect, modify, and distribute under a specified license) personal-agent project that went viral earlier this year, can already send and receive messages through a Twilio phone number or Messaging Service. Text may be the interface humans know best, but voice arguably carries more potential. Some AI startups are building on Twilio Programmable Voice for cases like this: an English-speaking user wants a room at a Tokyo boutique hotel that takes no online bookings; the agent places the call, asks about availability in Japanese, completes the reservation, and reports back in English on check-in, breakfast, and a late arrival.³ If a Jarvis-class assistant eventually arrives, it may well run on a stack like this one. On its most recent earnings call, Twilio disclosed that voice growth accelerated above 20% year-over-year, and we would argue the trend is only beginning.³

From Possession-like to Permission-Like SaaS

If the core value of possession-like SaaS is sharing an entry point, an interface, a user relationship, or a vertical workflow with human users, the second category's value lies elsewhere: controlling what an agent is allowed to do, what it can access, whom it may act on behalf of, and what verifiable record it leaves behind. This is the group we think is most likely to become the binding constraint (the bottleneck) as agents proliferate.

A few examples.

Cloudflare (1.65% weight in AGIX as of 8/31/2026*) could control network permissions. On its Q2 2026 call, the internet infrastructure company noted that non-human traffic had crossed half of all network traffic for the first time, and suggested that within five years, non-human traffic could be as much as 1,000x human traffic.⁵ We believe Cloudflare sits at the center of that structural shift, and through offerings like pay-per-crawl (a model in which an AI crawler, bot, or automated system pays a website or network provider for permission to access, retrieve, or process its content), it has begun charging for those requests. The implication is that in the agent era, the network may no longer be merely a transport layer; it becomes the permission layer that sits in front of action. What resources an agent may reach, under what identity it calls them, and at what frequency could all require Cloudflare's authorization, and Cloudflare could potentially extract value at that gate.

Snowflake (1.56% weight in AGIX as of 8/31/2026*) could control context and action permissions. Snowflake is a fully managed cloud platform for turning fragmented enterprise data into a shared, scalable foundation for analytics, applications, and AI. For enterprises, every AI-enabled workflow needs a governed, unified foundation of data and context. That foundation has real gravitational pull: once AI unlocks the value locked inside it, companies tend to migrate more and more data onto the platform, which in turn may accelerate the cloud data warehouse's growth. And a context platform is more than a data platform; over the long run, we believe data gravity may exceed model gravity. Management recently noted on the latest call that many customers are keenly interested in post-training open-source models on their own data inside Snowflake, and that this trend is strengthening.⁶

Datadog (1.22% weight in AGIX as of 8/31/2026*) could own agent reliability. In practice, no enterprise lets a model modify production simply because it can call tools. Real authority is granted only on top of continuous observability: the ability to see, in real time, what a system is doing and why. Datadog provides a unified software platform for monitoring the performance, reliability, and security of modern cloud-based applications and infrastructure. The company's agent console offers visibility into how AI agents are used, what they cost, and how well they perform. In a recent conference appearance, the CFO argued that model providers, database providers, vertical players, and GPU providers are all becoming tooling companies and infrastructure for the whole digital economy, and that they choose Datadog to monitor it.⁷ The deeper point is that agents seem unlikely to reduce the complexity of software systems. They may substantially increase the complexity of both producing and running software. If so, observability could be promoted from an ops tool to the feedback system of the agent economy.

Back to the Saaspocalypse

Recall what bearish investors actually worried about a few months ago. First, that AI platforms would commoditize software, letting ordinary users spin up their own tools on demand and discard them just as quickly. Second, that AI-native entrants would intensify competition across the industry, producing a wave of challengers. Third, that fewer software engineers would undermine the per-seat pricing model.

Set those against our view of what agents fundamentally are. Where AI can close the loop and iterate on its own capability in highly verifiable domains, coding being the clearest example, we would expect whole workflows to be abstracted away and, eventually, to disappear; software engineers may no longer define themselves primarily by writing code. But most knowledge work happens in open environments, where long-horizon tasks demand countless injections of human judgment.

For that reason, we believe the main line for a long time to come is this: keep agents and humans co-located inside possession-like SaaS, decompose long-horizon work into a great many small agentic automations, and steadily strengthen the capability and reliability of those agents through permission-like SaaS. AI ETF AGIX has positioned across both categories in an effort to capture the excess return a software renaissance could deliver.

---

Holdings are subject to change. *Holdings Data from Bloomberg as of 8/31/2026.

For AGIX standard performance, top 10 holdings, risks, and other fund information, please click here.

Citations:

1. Data from Bloomberg as of 8/31/2026.
2. SA Transcripts, "Salesforce, Inc. (CRM) Discusses AI Monetization, Product Momentum and Platform Integration Transcript," as of 9/1/2026.
3. Twilio, "Genspark scales 'Call for Me' AI agent globally with Twilio Voice," retrieved 8/31/2026.
4. SA Transcripts, "Twilio Inc. (TWLO) Q2 2026 Earnings Call Transcript," as of 8/7/2026.
5. SA Transcripts, "Cloudflare, Inc. (NET) Q2 2026 Earnings Call Transcript," as of 8/6/2026.
6. SA Transcripts, "Snowflake Inc. (SNOW) Q2 2027 Earnings Call Transcript," as of 9/2/2026.
7. SA Transcripts, "Datadog, Inc. (DDOG) Presents at Canaccord Genuity's 46th Annual Growth Conference Transcript," as of 8/12/2026.

Definitions:

Dow Jones U.S. Software Index (DJUSSW index): A float-adjusted, market-capitalization-weighted benchmark designed to measure the performance of U.S. companies classified in the software industry.

Drawdown: The percentage decline in an investment or index from its prior peak to its subsequent low point.

AI agent / agentic AI: An AI system designed to pursue a goal by using information, tools, software, and workflows with some degree of autonomy rather than only responding to individual prompts.

Artificial general intelligence (AGI): A hypothetical form of AI capable of learning, reasoning, and performing a broad range of intellectual tasks at a human-like general level.

Agent environment: A software application or workspace in which AI agents and human users can access context, collaborate, and complete tasks.

Agent stack: The collection of software, data, communications, security, payment, and infrastructure tools that allow AI agents to operate in real-world workflows.

Software as a service (SaaS): Software delivered online, typically through a subscription, rather than installed and maintained directly on a user's computer or company servers.

Annual recurring revenue (ARR): The annualized value of recurring subscription revenue, commonly used by software companies to indicate the scale of their subscription business.

Run-rate: An annualized estimate based on a company's current revenue, sales, or operating pace, rather than a full year of actual results.

Customer relationship management (CRM): Software used to manage customer data, sales opportunities, marketing activity, customer service, and business relationships.

Headless: A software architecture in which the data and business logic are separated from the main user interface, allowing those capabilities to be accessed through other applications or custom front ends.

Open source: Software whose underlying code is publicly available for users to inspect, modify, and distribute under a specified license.

Pay-per-crawl: A model in which an AI crawler, bot, or automated system pays a website or network provider for permission to access, retrieve, or process its content.

Graphics processing unit (GPU): A specialized computer chip originally designed for graphics that is now widely used to train and run AI models because it can perform many calculations simultaneously.

Per-seat pricing: A software-pricing model in which a customer pays based on the number of individual users granted access.