Today's AI agents are wrappers around interfaces built for humans. They click buttons. They parse screenshots. They re-read the same content from chat history every step. This is slow, expensive, and brittle. That's how almost every production agent in 2026 works.
We didn't get here because anyone designed it this way. We got here because the model came first. Models are spectacular at processing whatever you put in front of them. So we put human interfaces in front of them, and it worked well enough to demo. By the time anyone noticed that costs scale quadratically with the work, where every step drags the entire conversation history with it, there were already a million LangChain tutorials saying this was fine.
It is not fine.
The desk metaphor
Imagine you've hired a freelance researcher to write you a report. Two ways to manage them.
Way A: you keep them on a phone call the whole time. Every time they want to remember something, they read it back to you aloud. Every time they pick up a new source, they read it aloud again. By hour three the call is enormous, every minute costs a fortune, and they've half-forgotten what they decided in hour one.
Way B: they have a desk. They put sources on the desk. They write notes in a notebook. When they want to refer to source three, they pick it up off the desk. They don't read it aloud again. The desk is still there tomorrow. If they get sick, a colleague sits at the same desk and continues exactly where they left off.
Plynf is the desk, the notebook, and the filing cabinet for AI agents.
What that means in practice
State lives in the workspace, referenced by key. Each reasoning step gets a focused, small prompt. The token bill stays linear in actual reasoning work, not quadratic in conversation history. Tools route through one gateway, cached, audited, cost-capped, OAuth-handled. Instead of five bespoke integrations. Multi-agent coordination is structured messages on durable channels, not walls of pasted text. Workflows checkpoint. Crashes resume. Audits are tamper-evident.
None of this is novel infrastructure individually. What's novel is putting it all under one coherent API designed from day one for agents. Not retrofitted from a human UI, not assembled from five tools that don't know about each other.
Why Apache 2.0
Because the alternative, a proprietary platform with a "community edition," is the failure mode of every previous developer-infrastructure category. Vercel, Stripe, Resend, Linear: all built their commercial business on top of code that anyone could read, run, and replace if they got too greedy. That accountability is the entire point.
We will charge for the things teams genuinely want us to run. Hosted, multi-tenant, multi-region, SOC 2. We will not charge for the things teams should have for free: the runtime, the SDKs, the documentation, the right to fork.
The bet
In two years, every team running production agents will have some form of this layer. The question is whether they bought it, built it, or rented it from someone they regret. We are building the one they should buy. Open enough to be trustworthy, focused enough to be useful, hardened enough to be enterprise-grade.
If you are building agents in production, we'd like to hear from you.