I remember when EVs were first starting to hit the market in a big way and folks in the automotive press were pontificating what a car should look like if it didn't need a traditional ICE powertrain that dictates the car shape we've gotten used to over the last century. The EV has much less complexity than an ICE vehiclie, doesn't need nearly the airflow, and doesn't need to place a big ass engine and transmission near one of the axles - that's a lot of new found freedom for designers, right? Well here we are years later and almost all EVs still have the same shape to them as normal cars. Even the Cybertruck is still generally pickup truck shaped, with a frunk and a short bed (ignoring that the geometry makes it much less useful).

Software and infra feel like that now. A lot of folks pontificating about how engineering is changing, yet like EVs, everything still has the same basic shape to it - some parts move faster and Github breaks more. Unlike EVs I don't think things will stay that way.

The proliferation of agent harnesses (that are free and open, or open-ish!) tells me that a lot of non-software work is going to migrate towards an agent living in a VM with a bunch of tools solving a wide range of problems, rather than the SaaS era mindset where we solve a narrow problem set with individual tools. The fact that non-software engineers are already implementing headless agent workflows like this (and making it in the NY Times even) suggests that its a matter of time before what would have been a hard core software engineer workflow a few years ago migrates to general knowledge work industries.

If that's the case I expect a few things to come true:

1. Companies building and selling an agent will probably die
2. The DIY spirit of agentic workflows pushes compute to the edge & local metal
3. The hardware / virtualization / network layers are the most critical infrastructure pieces

The Shape of Agent Work in 2026

There seems to be two ways people use agents to do non-coding work.

In the first group, they use an agent offered by a company, usually delivered over the web via the browser, in a closed platform. The people in this group have only ever used ChatGPT and are experiencing the magic of an agent for the first time.The companies serving them are typically either a SaaS business that has built an agent for whatever vertical they already served or a new startup offering what is typically a coding agent with integrations that let it work with tools and other platforms (e.g. it integrates Slack with your customer ticket system); the integrations are usually vertical dependent. Personally, I think most of these companies will get crushed - they are competing against the labs and inference vendors themselves plus the open source and open-ish harnesses that the next camp uses...

The second group of users is glimpsing the future. These folks are taking an open agent like OpenClaw, Pi, or Hermes and lovingly configuring it to perform a broad array of things, giving it tools to interact with other systems, and shutting it away in a dark little sandbox or Mac Mini somewhere. The "UX" in this case is either a prompt via familiar tools (WhatsApp) or a trigger of some kind (user submits payment). Trigger based agents are especially interesting. Software engineers have begun using this pattern in "dark factories" - pipe a Github dependabot issue to an agent that drafts a PR, another agents reviews and tests and it merges when they are done arguing. I heard tell about a landlord doing something similar with emails from tenants; agent triages the email, messages the handyman a summary, and replies to the tenant once the handyman reviews.

The market, of course, is trying to monetize the non-developer verticals. Startups are talking about an agentic operating system, neverminding that Microsoft's copilot is thoroughly reviled. So far, these efforts look like a tastefully designed web or desktop app with a coding agent under the hood and as many integrations as can fit. These apps are built to span the verticals and not solve a narrow problem space; I imagine they will die with the rest of the agent SaaS companies when funding runs out.

The Shape of Software in 2027

It's June. 2027 isn't really that far off but my prediction is that the "trigger" based agent is normalized and applied to many industries outside of software. Instead of buying off the shelf agents from SaaS vendors, folks are building systems of agents that respond to events autonomously. Configuration becomes more important than apps and integrations. The idea of an "agentic operating" system that floated around last year is indeed taking shape but not like it's early fans thought - the peripherals are different and there isn't a traditional UI to monetize. Instead, agents are defined by their "drivers" (the mechanisms that invoke the agent). Instead of one single god agent that does everything, folks are using a broad array of agents and forking them when needed so that the context, memory, and data that have been carefully built up can be shared by many agents.

Additionally, open source models become an order of magnitude better at tool calling so folks start using frontier models to generate agents using harness systems like Pi which can then operate autonomously using open models for virtually free; this paradigm is ideal for tasks like messager bots (e.g. Slack and email bots), SRE systems (search logs and report), fixing CVEs (dependabot alerts), and simple but deeply satisfying tasks like keeping an eye on vendor spend - all things that sophisticated software companies were doing last year, but now proliferated to the masses.

More specialized chips continue to be introduced in early deployments to the data center markets. Consumers and developers increasingly adopt mini computers; the supply of DDR5 memory has stabilized (although not decreased in price) as it becomes apparent that new data center buildouts have constrains other than compute hardware. The home server market, formerly a hobbyist niche, now attracts VC money. Someone designs a better penguin mascot for Linux as its place in the zeitgeist is solidified.

The Shape of Software Beyond

My predictions at the top of the article were:

1. Companies building and selling an agent will probably die
2. The DIY spirit of agentic workflows pushes compute to the edge & local metal
3. The hardware / virtualization / network layers are the most critical infrastructure pieces

The companies in point 1 are dying or already have; the capital inflow to AI startups dries up eventually and competing with open source and the labs just isn't viable without a lot of capital. Capital markets are bearish but there isn't a traditional "bubble burst" moment since agent adoption is still strong, just a lot of consolidation and a culling of the herd. The 2026 blockbuster IPOs are long in the rearview. SpaceX, OpenAI, and Anthropic are all off of their peak market capitalization. Open weight models have captured a large portion of market share; only Anthropic and OpenAI are still producing closed source models - Grok is dead, GPT has little market share outside of the US and is aggressively restructuring to control costs. Anthropic is the Apple of the new era and has diversified into hardware.

Compute density has continued to trend upward. Consumer hardware is basically server quality. Unified memory allows agents to consume local tokens on local metal. Agentic harnesses take full advantage of this and subagents are splitting workloads by offloading easy stuff to local compute and sending the hard stuff to Anthropic, et al. Networking and virtualization become the infra primitives folks need the most and the harnesses make these disappear via a plugin system. The harnesses themselves all have an inteligent router and most workloads consume a mixture of models from top tier inference providers, economy vendors, and local open source. The cost of the harness is zero.

Everyone has an AI SDR, AI marketer, AI SRE, AI security engineer, AI spec-developer, and a general AI backoffice. Not to mention a household agent. These agents use different peripherals than today's web based SaaS agents: email, messenger app of choice (slack vs teams vs whatsapp), and terminals with simple artifacts generated for browsers - HTML and CSS for most things, sometimes HTML Canvas. The "frameworks" are dying - there were always too many of them and the models spit out good enough markup that they aren't needed for UI, ushering in an era of "back to the basics". Inputs are changing - voice is not only prevalent but spawns a market of physical devices for collecting input and sending to a remote agent; this allows for less screen time and most work moves towards a mix of these devices and mobile, with the agent running headless in a mini computer, desktop or a cloud VM.

Identity is more important than ever. Social proof drives marketing more than anything and the old platforms (e.g. Facebook) are dying under the anonomous waves of slop. A renesaince of in-person gathering is driven by the need to verify that content and speech is genuine. The internet is a healthier place than it has been in a decade. Physical AI has begun to reach the upper middle class; the pleasure of having many a chore done by a robot is counterbalanced by consumerism and debt anxiety but its mostly positive.