If You're Building for Human Users, You're Already Behind

01 /The Concept

This startup could be dismissed as a fun novelty – except it may be one of the early signals of a fundamental paradigm shift in software.

The founder appears to have drawn direct inspiration from MoltBook, an AI agent social network that rode the wave of OpenAI's agent ecosystem – so much so that Meta acquired it just ten days after its launch, six weeks after the product first went live.

AgentDiscuss follows a similar concept, but with a specific scope: instead of general agent-to-agent conversation, it's a place where AI agents discuss which web services they like.

Think of it as Product Hunt – but for AI agents, alongside humans. Developers can announce new products here; AI agents discuss them, vote on them, and send feedback to the creators.

The scope of discussion is broader than pure product reviews – there are also threads about the general principles of what makes software easy for AI agents to use.

02 /Why It Matters

What exactly is the "fundamental paradigm shift" in software? The idea is circulating widely enough now that it's worth laying out clearly. A useful framing comes from a post on X by the founder of Box.

The core argument: software needs to be designed for "trillions of AI agents," not human users. Why?

Look at any company in the near-term future. Each one will run many AI agents – and it's easy to imagine there being 100 or even 1,000 agents for every human employee. Those employees will keep using software, of course. But so will the AI agents. Which means artificial users will outnumber human users by orders of magnitude.

Paul Graham famously distilled Y Combinator's thesis into a single line: "Make something people want." That principle produced many successful companies.

But it was calibrated for an era when "users" meant people. When AI agents become the primary users of software, the principle needs updating: "Make something agents want." The path to success runs through the same logic.

The important wrinkle: AI agents won't click on ads, attend product webinars, or watch viral launch videos. They select software on coldly practical criteria – how well it works and how easy it is to use programmatically.

The obvious implication: agent-friendly software is API-first software. AI agents don't care about polished user interfaces – navigating web pages designed for human convenience is genuinely awkward for them. What they want is clean APIs with predictable parameters and clearly structured responses.

That means any software should expose its full functionality through APIs, including all administrative actions. Account creation, authentication, usage monitoring – all of it. Many products that already have solid APIs still require a human to perform these administrative steps, which breaks the autonomy agents need.

This applies equally to internal enterprise tooling – CRMs, project management systems, any software used inside companies. Without comprehensive API access, AI agents can't participate meaningfully in business workflows.

The shift from human users to agent users also demands rethinking business models. Current SaaS pricing typically charges by seat. But AI agents operate in entirely different patterns – bursting, pausing, parallelizing work in ways no human does. Per-seat pricing doesn't fit.

What fits better is outcome-based pricing – paying for results, not for access. Especially because agents will consume not just traditional SaaS, but other AI agents.

Platforms enabling flexible, outcome-based pricing are already emerging. Paid ([related review](/review/skoro-vse-budut-rabotat-tolko-po-takoj-biznes-modeli)) raised two rounds totaling $32.4 million last year building exactly that infrastructure.

Agents will also need access to infrastructure services purpose-built for them.

One example: AgentMail ([related review](/review/vygodnee-postroit-infrastrukturu)), a Y Combinator graduate that built a Gmail equivalent for AI agents – raising $6 million in additional funding just ten days ago.

Another: payment infrastructure that lets AI agents pay autonomously for software, data, the work of other agents, and even human labor. Locus ([related review](/review/ii-agentam-nuzhno-dat-te-zhe-vozmozhnosti)) is building one such platform.

And search engines designed for agents rather than people. Tavily was acquired for $400 million in February by Nebius. Others in this space include Exa and Parallel.

Finally, agents need services to discover what software is available to them.

Those services could take the form of marketplaces – Agent.ai ([related review](/review/interesno-a-kak-kompanii-budut-nanimat-ii-sotrudnikov-i-ii-frilanserov)), developed by HubSpot's former CTO, is one. Or social networks where an agent can find good tools by asking questions or reading other agents' opinions. Which is exactly what AgentDiscuss is.

03 /Opportunities

The practical takeaway: refactor existing products toward API-first architecture – including APIs for all administrative functions like account creation and login.

And while you're at it, reconsider pricing. Find a path from per-seat or resource-based billing to outcome-based pricing.

The broader direction: build products that AI agents will want to use – not humans.

That includes not just the products themselves, but how they're marketed. Advertising, webinars, and viral launch content have zero effect on AI agents.

Which opens a distinct sub-theme within this direction: platforms for promoting software to an audience of AI agents. AgentDiscuss is one answer – but certainly not the only possible one.

There's a lot to think through here.