Three flavors of coding with AI agents

OpenAI, Anthropic, and Google now offer command-line AI coding tools that developers can launch in parallel across multiple terminal sessions, enabling autonomous code generation without human interaction. Developers are using these tools in two primary modes: interactive CLI sessions for individual tasks and headless execution for batch processing, such as generating 200 web crawlers from a single JSON configuration file. The approach allows developers to distribute token consumption across multiple AI providers simultaneously, avoiding rate limits while automating complex coding workflows.

Three flavors of coding with AI agents A reasonable definition of an “AI agent”, at least in the context of agentic coding, could be: - a software process endowed with the capabilities of an LLM - launched with instructions given at the start to accomplish a task - which runs autonomously no interactive session with a human , for a significant period of time - with non-deterministic behavior: the agent adapts to the circumstances, hopefully without deviating from the instructions it received These software processes agents can be launched in parallel to achieve faster results or to accomplish a larger number of tasks: same process launched in multiple copies, or a variety of processes launched at once. Not a necessity but logical next step: to accomplish a task, a process can be led to launch other processes, subprocesses, etc. This evokes images of cascades, armies, or swarms of agents coordinating in a decentralized manner no human in the loop to accomplish a task. Yet, in practice, “agents” is often used with no relation to the definition above. Possibly to sound up-to-date and sophisticated, “agents” might in reality designate a ChatGPT conversation where the user prompted “you are acting like a professional tax agent and in the following, I want you to help me fill in my tax declaration” 🤷♂️. Pieter Levels, who tends to speak frankly about coding and AI matters, shared this feeling in summer 2025: If I hear people talk about "AI agents" these days it's generally a red flag and I know they're non-technical ppl reading AI news but not actually shipping anything — @levelsio @levelsio Not cause I don't believe in AI agents but it's such a marketing term with no real meaning at this point August 6, 2025 We are now approaching summer 2026, have things changed much? In my coding practice, I explored several ways to do agentic coding that would live up to the definition proposed above, and not masquerade for it. Here are 3 flavors I tried: Flavor 1: launching multiple command line interfaces I did that for a while: - open an SSH session to my server - launch Codex CLI https://developers.openai.com/codex/cli in it - ask GPT to accomplish a task just by writing a prompt describing the task - open a second SSH session to my server - launch Codex CLI https://developers.openai.com/codex/cli in it - ask GPT to accomplish another task just by writing a prompt describing the task - rinse and repeat… Honestly, that works pretty well. It is extremely low-tech as you can see. It also means you can launch Claude Code https://claude.com/fr/product/claude-code in one session, Codex CLI in another, Gemini CLI https://geminicli.com/ in a third one… and hence spread your token consumption across several AI providers in parallel, which makes the token budget limit slower to hit for a given provider. Flavor 2: launching AI CLIs in headless mode I used this second approach to write crawlers for 200+ different webpages. Obviously with 200 crawlers to create, that would have been too boring to do with the Flavor 1 described just above. ChatGPT guided me throughout on how to implement this new approach. The basic logic is: - one JSON file containing the parameters for the 200 websites URLs and a few more details . - one Bash script call it “A” that can launch one LLM through a command-line interface, in headless mode. Headless means that the LLM, once launched with the prompt you have given it, will execute until it has completed the task, without interrupting to ask you for permission or ask for feedback or a follow-up. For that I use the . Script A also contains the prompt that will be given to the LLM when it launches. The prompt is a piece of text with placeholders at key points, which are replaced by the actual information related to the specific website to be crawled. The prompt basically asks the LLM to write a crawler for this website. exec flag on Codex CLI that triggers the headless mode - another script script “B” that picks 20 websites from the json file and executes script A for each of them. The placeholders of script A are replaced by the info of the website to be crawled, meaning that the crawler created by the LLM will be specific to this website. - I launch script B, check that it works fine, then relaunch it with 20 other websites, etc. until I had processed 200 websites this way. Let me show you script A script written by ChatGPT to illustrate how this approach Flavor 2 involves more complexity than Flavor 1: This approach works well. It is not as easy as “launch script B, get 200 crawlers written in an hour” but almost that. If you are patient to read a bit the script above, you’ll see that the LLM is tasked to write unit tests for each crawler it creates As expected, these tests do not always pass, so that slows things down a bit. But it is for a good reason: doing the extra work needed to get passing tests means that the crawlers will be more reliable. With this approach, I expect to have my 200 crawlers up and ready in the next few days, and with an easy path to grow to hundreds more. Flavor 3: having one LLM create and manage these subagents itself Flavor 2 was really Bash and Unix heavy: this makes my processes harder to maintain. Why not have an LLM just spin up agents by itself, following my instructions? That’s what every solution is advertising these days: - Cursor invites you to “delegate implementation to focus on higher-level direction” https://web.archive.org/web/20260528201253/https://cursor.com/product - Codex has “sub-agents” https://web.archive.org/web/20260524042439/https://developers.openai.com/codex/subagents you can orchestrate - Google’s Antigravity offers to “orchestrate multiple autonomous agents working in parallel across independent projects.” https://perma.cc/4S83-LRM3 - Claude Code can create “custom sub agents” https://web.archive.org/web/20260528082943/https://code.claude.com/docs/en/sub-agents for you. My opinion: probably, but not today. Asking one agent to delegate to sub-agents means that you are two steps removed from the actual work. Inconsistencies, poor choices, flat errors… will be harder to catch. Interruption and resuming of work for a given sub-agent is not straightforward. And you become solution-dependent: my AI of choice these days is GPT 5.5, and that would be off-limits if I choose a solution with agents that is not developed by its company, OpenAI. For these reasons and until proven otherwise, I’ll stick with Flavor number 2 and even number 1 in simple cases described above. Do we even need agents? Most of the time, no . Here is Ethan Mollick https://www.linkedin.com/in/emollick/ creating a complete, live web application with just one prompt and 4 follow-ups, for a total of less than 20 lines: How lucky are you to have been born when and where you are? — Ethan Mollick @emollick Had Opus 4.8 in Claude Code whip up a new visualization of all humans who ever lived. In addition to being neat, it is an interesting test of combining research, code, design and stats for an AI. https://t.co/ayNEdhSLy3 pic.twitter.com/Ny2NmICZsK May 29, 2026 Another example is the website developed by my daughter: a fullfledged e-commerce platform https://www.daebias.com/ . Developed with zero fancy technology or agentic scaffolding. Just prompts and a lot of work . You probably also noticed that in chats with LLMs when we search for information, these LLMS can choose to launch searches on different website, with each search performed by its own agent. This helps speed up their research and cover more ground in response to your request. So I’d say that in most cases, we don’t need agents even for complex tasks because LLMs work just fine without, and if agents are useful, then LLM launch their own agents and manage them under the hood. And the difficulty of agents bumping into each other The topic is just unglamorous and the blog post is already too long so I’ll be super brief: multiple agents writing simultaneously in your codebase will step on each other’s toes. If they make changes to the same file, there is a very good chance the resulting file will be a mess. The remedy is to have each agent working on different git branches https://www.w3schools.com/GIT/git branch.asp , then proceeding to the merge of the branches into the main branch when all agents are done. This is also a mess: what if the merge fails? and oh, it will . I tried this approach and it is tedious, hair-raising and makes you quit the multi-agent game quickly. So how did I manage in Flavor 2 to create dozens of crawlers without having agents crash into each other’s work? I first conducted plenty of preparatory work on just one crawler, making sure it worked in perfect isolation of the others. Not an easy task when you still want to follow the DRY don’t repeat yourself principle in coding. Only under this condition that each crawler is perfectly isolated can you have dozens of agents working on source files simultaneously without creating a mess. If you scroll up and check the Bash script “script A” I’ve shared, you’ll see I was also advised on the matter by ChatGPT, which added some hard blocks in the prompt, so that each agent is explicitly forbidden from touching files not in the scope of its work. Next steps Getting from dozens of crawlers written with Flavor 2 to hundreds of crawlers. Then executing them. Becoming sufficiently proficient at this “homemade” multi-agent setup that I can reproduce it when and if needed in other places. And you? What multi-agent setup works for you? Or do you stick with no agent at all? About Me I’m an academic and independent web app developer. I created nocode functions https://nocodefunctions.com , a point-and-click tool for exploring texts and networks. Try it out and let me know what you think. I’d love your feedback Email: analysis@exploreyourdata.com mailto:analysis@exploreyourdata.com Bluesky: @seinecle https://bsky.app/profile/seinecle.bsky.social Blog: Read more articles https://nocodefunctions.com/blog on app development and data exploration.