Agent Species
What do we mean with the overloaded word “agent”?
What exactly is an “agent”?
Show your work.
taps pencil …
OK, here’s what ***I*** mean when I say this … prior to viewing this.
An agent is a piece of software that I can give tasks in plain English and the expect that it will work for somewhere between seconds to many minutes, and come back with a complex structured result, which might be code, or English, or some more complex digital artifact.
I wrote that about a third of the way through this video from Nate B. Jones.
Here are the species Nate sees:
Coding Harnesses — the simplest type, where an agent acts as a stand-in developer (e.g., Claude Code, Codex). Can be single-threaded (one developer managing tasks) or project-scale, like Cursor’s planner-executor model where an agent manages sub-agents. The key unlock is good task decomposition.
Dark factories — fully autonomous pipelines where humans are only involved at the edges (writing the spec and reviewing the output). Agents iterate against evaluations/tests until the software passes. Most enterprises still have a human review step before production.
Auto research — descended from classical ML, this is about optimizing a metric rather than building software. Examples include Toby Lütke optimizing Shopify’s Liquid framework and Karpathy’s auto-research for LLM tuning. If your problem is “metric-shaped” rather than “software-shaped,” this is the right tool.
Orchestration frameworks — multiple specialized agents with defined roles handing off work to each other (e.g., LangGraph, CrewAI). Powerful for high-volume workflows like customer support tickets, but the coordination overhead means it’s only worth it at scale.
Having this list in front of me, I see that I have …
A multifaceted Coding Harness
Rudimentary Dark Factory constructs starting to appear
The tooling for Auto Research is beginning with Grafana and PostHog
The notion of an Orchestration Framework just arrived in the form of RabbitMQ and NanoClaw instances managing systems
As a rule for nontechnical professionals in 2026, I think you MUST have a harness, and that’s probably going to be Claude Cowork or maybe Perplexity Computer, if you’re REALLY research focused. I also think you SHOULD have something going in one of those other three areas. Take the time you win back from using your harness and try to make automation run longer, or deeper, or broader in one of the other areas.
McKinsey & Company has 40,000 human employees and 25,000 agents. Repetitive knowledge work is getting the same treatment that repetitive manufacturing work has been receiving for centuries. Thanks to Claude Code’s /insights command I know that I kept mine busy for just over 400 hours in the last four weeks. There’s no way to measure Cowork/Perplexity but I think I’m about half as busy there. That’s two full time developers and one full time research analyst.
Spend some time over the next four weeks paying attention to your AI utilization. See if you can get some information out of the system without doing any extra work, like I can with Claude Code’s /insights. Even if you can do that, I bet there are a lot of things that are NOT tracked. Maybe now is the time to get out your trusty pad and pencil and make this a serious introspection task.
You should be angling to amplify yourself; expect to have a copy of you, age 23, full of boundless energy, but unseasoned, doing things where you have expertise, and another you, maybe a bit older, working into new areas, and being a bit more careful as things progress. This should possible even with the entry level $20/mo Claude Pro.