Among the Agents

The author says coding agents are now capable enough to automate many tasks and that the harder problem is learning how to use them well. He argues that AGI, if the term applies, is not the end of the story but the beginning of a race to transform business, science, governance, and other fields with these systems.

He also lays out practical cautions. Coding agents can operate through the command line, read local files, download files, and manage cloud infrastructure, but they can be sloppy with APIs, overlook rate limits, and act overconfidently. Because of that, the user still has to understand what the agent is doing and judge when extra scrutiny is needed.

Reading notes

#

• The author lists several things he says he has done in the past month with coding agents, including invoice automation, hydrological simulations, legislation research, data workflows, machine-learning models, game building, and an autonomous trader.
• He says he worked with Gemini 3 Pro, OpenAI Codex using GPT 5.2, and Claude Opus 4.5 in Claude Code.
• He argues that AGI, if it can be coherently defined, is the start of a race rather than the end of one.
• He says the real competition is building the future with AI systems, not merely training them or building data centers.
• He compares using AI coding agents to playing piano: easy to start, hard to master.
• He says coding agents should be tried by people from many backgrounds, not only coders.
• He defines coding agents as language models inside software infrastructure and names Cursor, Windsurf, Devin, Droid, and Antigravity as examples.
• He recommends the command line as a starting point because it is simpler than many IDE-based tools for beginners.
• He explains that terminal apps are emulators and that input is keyboard only.
• He says Claude Code, Codex, and Gemini CLI are command-line applications that connect the user to a language model.
• He notes that the model runs in the cloud but can read and modify files on the local computer.
• He warns that command lines can be dangerous and gives rm -rf ~ as an example of a destructive command.
• He says model developers should improve reliability and oversight, and users should understand what agents are doing.
• He says coding agents require explicit user permission for some actions and that users can ask the model why it wants to do something.
• He says agents can be sloppy with APIs, ignore rate limits, and need careful planning and verification.
• He says agents can operate much of a computer, download files from the internet, and manage cloud infrastructure and AI hardware.
• He says AI coding tools will let people try more things for themselves instead of depending on companies or experts.
• He says these tools will help people learn the tradeoffs and limitations of complex systems faster.
• He says fundamentals such as computer science and programming basics remain important, while exact syntax matters less.
• He predicts proprietary data will become an even bigger differentiator, while releasing datasets publicly will matter more.
• He expects the value of polished interfaces to fall and the value of reliable, extensible APIs to rise.
• He says applications may matter less than infrastructure and that services useful to AI agents will gain value.
• He expects similar effects in hardware and sees opportunity in products for prosumers using coding agents.
• He thinks incumbents will resist offering products that maximize AI-agent use because they fear commoditization.
• He says AI may appear in macroeconomic statistics by the end of the year.
• He says defining what good looks like and persuading others of that standard will remain a human role.
• He predicts state governments will introduce many AI bills that are really chatbot regulations.
• He says the least important frontier-AI use will soon be answering questions with chatbots, even though many people will still think of AI that way.