What Coding Agents Actually Are

Edition 166 | March 16, 2026

Chipotle’s support agent helps us buy burritos and learn Python.

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• Musk reveals the Macrohard agent architecture

• Karpathy open-sources an overnight research agent

• LangChain's GTM agent converts 250% more leads

• Virgin Voyages scales to 1,500 AI agents in four months

• China's OpenClaw gold rush goes mainstream

…and more

🔍 SPOTLIGHT

Everyone keeps asking when the year of agents will come. It might have already arrived and we all didn’t notice.

A few weeks ago, SemiAnalysis published a piece called "Claude Code is the Inflection Point". If you're not familiar, SemiAnalysis is one of the most respected semiconductor research firms in the world. What caught my eye wasn't the coding stuff. It was what their non-coders were doing with it. Their datacenter model team now reviews hundreds of documents a week through Claude Code. Their supply chain team inspects bills of materials with thousands of line items. Their memory model team built real-time forecasting tools as spot market prices moved. Their technical staff spun up a live dashboard that runs nightly across nine different system types. Nobody asked these people to build these things. They just had problems, described them in plain language, and got working solutions.

It is, put simply, an agent that can code. Not just a coding agent if you are following.

I think most people hear "Claude Code" or "Codex" and immediately label it as “coding stuff." And I get why it seems alienating. You use it in a terminal or code editor. But here's the thing I've realized after months of using it: you're not coding. You're chatting with a virtual agent. You describe what you want, and the agent figures out how to make it happen. The code is just the language it uses to take action on your computer. Behind every button you click on Amazon, every Slack message you send, every Google Sheet formula, it's all just code running underneath. These agents speak that language natively. You describe outcomes. The agent handles the rest.

And that's the reframe I want you to sit with. The skill isn't coding. It's orchestration. You could call it management. Knowing what problem to hand off, what context to provide, what tools to equip. You can give Claude Code access to your Gmail and have it run lead outreach. Hand it a messy spreadsheet and get back a polished report. Give it skills (reusable instructions and tools) and it becomes a repeatable workflow machine. The real art is in the setup, not the syntax.

SemiAnalysis frames the bigger picture perfectly. All information work follows the same loop: READ (ingest unstructured information), THINK (apply domain knowledge), WRITE (produce structured output), VERIFY (check against standards). That's what a coding agent does for software. But it's also what an analyst does with earnings reports, what a consultant does with market research, what a paralegal does with contracts. The workflow is identical. Agents just execute it faster.

And the economics are hard to ignore. A Claude Pro subscription runs $20/month. Max is $200. The median US knowledge worker costs $350-500 a day fully loaded. An agent that handles even a fraction of their daily workflow is a 10-30x ROI. 64% of developers now use AI agents per SonarSource's 2026 State of Code report, up from just 31% a year ago. 4% of all GitHub public commits are currently authored by Claude Code, and SemiAnalysis projects that number hits 20%+ by end of 2026. We are early on the adoption curve, and this is just the beachhead.

And yes, there are still gaps. Memory isn't perfect, context windows have limits, you still need to set things up with intentionality, prompt injection, etc. But these “coding” agents are doing real work. Today. Coding is just where these agents landed first because the digital world runs on code. The $15 trillion information work economy is next. And the people who are building muscle memory with these tools right now, even if they've never opened a terminal before, are going to have an enormous advantage as agents get better.

If you haven't tried Claude Code or Codex yet, start this week. You don't need to be a developer. You just need a problem to solve. And we all have plenty.

As always, keep learning and building!

—AP