I mean "someone" as in the author of the template (to use your phrasing), yes.

> What are the specific limitations of agents? What can they do?

What each agent does, not matter where they sit in the network layers, is the same; they request information from their clients, they make some decisions and then they contract other agents. In this system, everything is a (compiling) supplier. A developer creates an agent to build on and inherit the knowledge of its suppliers, adding a little knowledge himself in the process, and ensuring that his agent returns compiled code.

Have you had a chance to check out this video (https://www.youtube.com/watch?v=qTl-V58cG1w)? It sheds a bit more light on how the service an agent performs is that agent doing its own part in a distributed compiler.

A developer can only build an agent if its suppliers exist, and if the compile-time interface protocols have been defined. Are they the sorts of limitations you meant?

> How are they written?

This system bootstrapped itself (and achieved PoC at the same time) back in 2013. Agents are built by agents. There is a special agent that sits on top of this hierarchy of agents that is designed to specifically capture a developer's requirements for their agent (what it should ask its client, what decisions it should make, and who it should contract) and translate these requirements to contracts to agents which will ultimately build that developer his agent. (The recursion is a killer sometimes, isn't it?)

This special agent that sits on top of the hierarchy is called the 'Agent-builder' and will feel to a developer like an IDE or visual programming language (designed specifically to design agents).