Agentic Coding Recommendations
There is currently an explosion of people sharing their experiences with agentic coding. After my last two posts on the topic, I received quite a few questions about my own practices. So, here goes nothing.
Preface For all intents and purposes, here's what I do: I predominently use Claude Code with the cheaper Max subscription for $100 a month . That works well for several reasons: I exclusively use the cheaper Sonnet model. It's perfectly adequate for my needs, and in fact, I prefer its outputs over the more expensive Opus model.
I optimize my tool usage to be token efficient. I avoid screenshots and browser interactions wherever possible. More on that later. My general workflow involves assigning a job to an agent (which effectively has full permissions) and then waiting for it to complete the task. I rarely interrupt it, unless it's a small task. Consequently, the role of the IDE — and the role of AI in the IDE — is greatly diminished; I mostly use it for final edits. This approach has even revived my usage of Vim, which lacks AI integration. One caveat: I expect this blog post to age very poorly. The pace of innovation here is insane; what was true a month ago barely holds true today. That's why I'm sticking to concepts I believe have staying power. If you want to a small session of me working on an Open Source library with it, I have a recording you can watch.
The Basics I disable all permission checks. Which basically means I run claude --dangerously-skip-permissions . More specifically I have an alias called claude-yolo set up. Now you can call that irresponsible and there are definitely risks with it, but you can manage those risks with moving your dev env into docker. I will however say that if you can watch it do its thing a bit, it even works surprisingly well without dockerizing. YMMV. MCP. This is a term you cannot avoid. It basically is a standardized protocol to give agents access to more tools. Honestly: at this point I barely use it, but I do use it. The reason I barely use it is because Claude Code is very capable of just running regular tools. So MCP for me is really only needed if I need to give Claude access to something that finds too hard to use otherwise. A good example for this is the playwright-mcp for browser automation. I use it because I haven't found anything better yet. But for instance when I want my agent to poke around in my database, I just uses whatever it finds to be available. In my case it loves to use psql and that's more than good enough. In general I really only start using MCP if the alternative is too unreliable. That's because MCP servers themselves are sometimes not super reliable and they are an extra thing that can go wrong. Trying to keep things very simple. My custom tools are normal scripts that it just runs.
Choice Of Language I've evaluated agent performance across different languages my workload, and if you can choose your language, I strongly recommend Go for new backend projects. Several factors strongly favor Go: Context system: Go provides a capable copy-on-write data bag that explicitly flows through the code execution path, similar to contextvars in Python or .NET's execution context. Its explicit nature greatly simplifies things for AI agents. If the agent needs to pass stuff to any call site, it knows how to do it.
Go provides a capable copy-on-write data bag that explicitly flows through the code execution path, similar to contextvars in Python or .NET's execution context. Its explicit nature greatly simplifies things for AI agents. If the agent needs to pass stuff to any call site, it knows how to do it. Test caching: Surprisingly crucial for efficient agentic loops. In Rust, agents sometimes fail because they misunderstand cargo test 's invocation syntax. In Go, tests run straightforwardly and incrementally, significantly enhancing the agentic workflow. It does not need to figure out which tests to run, go does.
Surprisingly crucial for efficient agentic loops. In Rust, agents sometimes fail because they misunderstand 's invocation syntax. In Go, tests run straightforwardly and incrementally, significantly enhancing the agentic workflow. It does not need to figure out which tests to run, go does. Go is sloppy: Rob Pike famously described Go as suitable for developers who aren't equipped to handle a complex language. Substitute “developers” with “agents,” and it perfectly captures why Go's simplicity benefits agentic coding.
Rob Pike famously described Go as suitable for developers who aren't equipped to handle a complex language. Substitute “developers” with “agents,” and it perfectly captures why Go's simplicity benefits agentic coding. Structural interfaces: interfaces in Go are structural. If a type has the methods an interface expects, then it conforms. This is incredibly easy for LLMs to “understand”. There is very little surprise for the agent.
interfaces in Go are structural. If a type has the methods an interface expects, then it conforms. This is incredibly easy for LLMs to “understand”. There is very little surprise for the agent. Go has low eco-system churn: Go's entire ecosystem embraces backwards compatiblity and explicit version moves. This greatly reduces the likelihood of AI generating outdated code — starkly contrasting JavaScript's fast-moving ecosystem for instance. For comparison, Python — my initial choice — often poses significant challenges. Agents struggle with Python's magic (eg: Pytest’s fixture injection) or complex runtime challenges (eg: wrong event loop when working with async), frequently producing incorrect code that even the agentic loop has challenges resolving. Python also has practical performance problems. I don't mean that it writes slow code, i mean that the agent loop is really slow. That's because the agent loves to spawn processes and test scripts, and it can take quite a while for the interpreter to boot up and initialize the entire application. On the frontend I settled on tailwind, react with tanstack's query and router as well as vite. I'm not amazingly happy with it, but I found it better than the alternatives. Tailwind and vite are great, no complaints there. Tanstack's file based router does not make me super happy. In parts because it likes to have dollar signs in the file names and those really like to confuse the agent. For instance it's quite common that it tries to edit $param.tsx but it edits the file .tsx instead because it to confused by shell interpolation. It's a minor thing, but a very annoying thing.
It's All About Speed Agentic coding's inefficiency largely arises from inference cost and suboptimal tool usage. Let me reiterate: quick, clear tool responses are vital. What we did not talk about yet is that some tools are "emergent," temporarily written by agents themselves. Quick compilation and execution significantly boost productivity of the agent. So how can we help it? With the right instructions it must be possible for the AI to create a new tool, by following existing conventions very quickly. This is necessary because you want the AI to write some new code, and run it. There is a big difference in the quality and speed of the flow, if that tool takes 3ms to run vs it compiles for 5 seconds and then needs another minute to boot and connect to database and kafka broker and 100 lines of nonsensical log output. If your stuff is indeed slow, then consider vibe-coding a demon that you can dynamically load stuff into. As an example Sentry takes too long to reload code and it takes too long to restart. To trial some agentic coding there my workaround was a module that watches a file system location and just imports and executes all python modules placed there, then writes the outputs into a log it can cat. That's not perfect, but it was a significant help for the agent to evaluate some basic code in the context of the application. Balancing log verbosity is crucial: informative yet concise logs optimize token usage and inference speed, avoiding unnecessary costs and rate limits. If you cannot find the balance, provide some easy to turn knobs for the AI to control. In an idea setup you get useful log output as a natural byproduct of the agent writing code. Getting observability from the first shot of code generation beats writing code, failing to run it and only then going back to a debug loop where debug information is added.
Stability and Copy/Paste Stable ecosystems are what you really want. LLMs are great with Go and they love to use Flask, because those are quite stable ecosystems with little churn. The same thing is true for your codebase. The AI likes to leave all kinds of breadcrumbs lying around when writing code that can turn into confusion later. For instance I have seen the agents leave useful comments about why it chose one path over another. If you nilly-willy let the AI upgrade libraries where some of those decisions no longer make sense, you now might have the AI continue making a now outdated pattern. In theory this should be the same for agents and humans, but the reality is that agents make upgrades so “cheap” that it's tempting to just let the AI do it and see if tests still pass. I do not find this to be a successful path at all. Be even more conservative about upgrades than before. Likewise with AI I strongly prefer more code generation over using more dependencies. I wrote about why you should write your own code before, but the more I work with agentic coding, the more I am convinced of this.
Write Simple Code Simple code significantly outperforms complex code in agentic contexts. I just recently wrote about ugly code and I think in the context of agents this is worth re-reading. Have the agent do “the dumbest possible thing that will work”. Prefer functions with clear, descriptive and longer than usual function names over classes.
with clear, descriptive and longer than usual function names over classes. Avoid inheritance and overly clever hacks.
Use plain SQL . I mean it. You get excellent SQL out of agents and they can match the SQL they write with the SQL logs. That beats them min-maxing your ORM's capabilities and getting lost in the SQL output in a log.
. I mean it. You get excellent SQL out of agents and they can match the SQL they write with the SQL logs. That beats them min-maxing your ORM's capabilities and getting lost in the SQL output in a log. Keep important checks local. You really want to make sure that permission checks are very clear to the AI, and that they are taking place where it AI can see it. Hiding permission checks in another file or some config file will amost guarantee you that the AI will forget to add permission checks in when adding new routes.
Make It Parallelizable Agents aren't exceptionally fast individually, but parallelization boosts overall efficiency. Find a way to manage shared states like the file system, databases, or Redis instances so that you can run more than one. Avoid them, or find a way to quickly segment stuff out. Your initial shared state is just the file system and a second check-out will do. But really I don't have an amazing solution yet. There are some good initial attempts. For instance one of the tools to watch is container-use. It's an MCP server that instructs Claude or other agents to run their experiments entirely in Docker. Then there are tools like Cursor's background agents and Codex which are moving this entire stuff into CI which will be interesting. So far, I don't this is working for me yet, but let's see again in a month.
Learn To Refactor Agentic coding alters refactoring priorities. Agents handle tasks effectively until project complexity surpasses some manageable thresholds. Too big here is defined by the total amount of stuff that it has to consider. So for instance you can vibe code your frontend together for a while, but eventually you reach the point where you absolutely need to tell it to make a component library. Why? Because if the total tailwind class mess is splitered across 50 files you will find it very hard to get the AI to make redesigns or extract components without major regressions. An agentic workflow encourages good code maintenance and refactoring at the right moment. You don't want to do it too early and you definitely do not want to do it too late.