The best survey of the agent economy isn't from Gartner or McKinsey — it's 334 engineers on Hacker News comparing notes on what they actually pay for.
The Summary
- A developer asked HN how people are using AI daily, and 334 responses paint a picture of the real agent economy: messy, pragmatic, rapidly evolving
- The thread reveals a split between "subscription stacks" (multiple paid tiers across providers) and "API minimalists" (pay-per-token power users)
- Key pattern: developers are treating AI tools like infrastructure now, not experiments — they're building multi-agent workflows and custom harnesses
- The conversation itself is the signal: what people are paying for shows where the value actually is
The Signal
The thread started simple: what AI tools do you pay for, and how do you use them? But scroll through 334 replies and you see something bigger than a product review thread. You see the shape of Web4 forming in real-time.
The most common setup: ChatGPT Plus for general work, Claude Pro for writing and reasoning, Cursor or Windsurf for coding. Multiple subscriptions aren't unusual anymore — they're table stakes. One developer listed five concurrent subscriptions totaling $150/month. Another mentioned maintaining two Claude accounts to work around rate limits. This isn't hobbyist behavior. This is people running AI like they run AWS.
"What people pay for reveals what actually works — and they're paying for everything."
The subscription layer is just the surface. Beneath it, a second tier of users is building custom harnesses and orchestration systems. They're using tools with names you won't see in TechCrunch: Aider for command-line AI pair programming, OpenRouter to route between models based on task type, locally hosted models for privacy-sensitive work. One developer described a setup where different agents handle different parts of the coding workflow — one for architecture, one for implementation, one for testing. Another built a custom harness that switches between GPT-4 and Claude mid-conversation based on the type of reasoning required.
This is the agent swarm pattern emerging in the wild, without fanfare. People aren't waiting for "agent platforms" to mature — they're building their own orchestration layers. The original poster listed specific tools: Munder Difflin, Gas Town, Ruflo. Most respondents had never heard of them. But many had built something similar themselves.
The API-first users are a different breed entirely. They're not paying $20/month for ChatGPT — they're running $500+ monthly bills through OpenAI's developer API, Anthropic's direct access, or aggregators like OpenRouter. One comment thread became a mini-tutorial on token economics: when to use which model based on cost per million tokens, quality thresholds, and latency requirements. These users think in compute, not features.
Key distinctions in the thread:
- Subscription users: Pay for convenience, UX, and avoiding rate limits
- API users: Pay for control, cost optimization, and ability to chain models
- Hybrid users: Subscriptions for daily work, APIs for production systems
What's striking is how fast the baseline shifted. A year ago, most HN users were experimenting with ChatGPT. Now they're debating orchestration patterns and complaining about Claude's context window limits like they're complaining about S3 pricing. AI tools crossed the chasm from "cool demo" to "critical infrastructure" without anyone declaring victory.
The thread also reveals the gaps. Several developers mentioned frustration with "broader agent platforms" — the VC-funded startups promising to orchestrate everything. The consensus: too rigid, too slow to adapt, too much lock-in. The builders want composable primitives, not opinionated platforms. They want to own the orchestration layer, not rent it.
"The real agentic setup isn't a product you subscribe to — it's a stack you compose."
One developer summed it up: "I'm paying for three coding assistants, two general LLM subscriptions, and running my own local model for anything sensitive. A year ago I had zero. Two years ago this would have sounded insane." That's the baseline now. Not for everyone, but for the people building the next layer of software.
The Implication
If you're building in the agent economy, this thread is a better market map than any analyst report. The money is flowing to tools that compose well, not tools that try to own the whole workflow. Developers are willing to pay significant monthly bills, but only if they can mix and match, build custom harnesses, and maintain control.
Watch what happens next: the subscription model will hit a ceiling as power users shift more spend to APIs and self-hosted models. The winner won't be the platform with the best UX — it will be whoever makes it easiest to orchestrate multiple specialized agents without lock-in. Web4 isn't one app. It's a stack you build yourself, agent by agent.