"If you are only using your hammer to hammer nails, you're doing it wrong" then goes on to explain how you should use agents.
I would've thought that following the initial argument and the progression to the latest trend we would've ended at use agents and write specs and these several currently popular MCPs.
I guess my rant is it to arrive at the point that no one knows what the "correct" way to use them is yet. A hammer has many uses.
I want to believe, and I promise I'm not trying to be a luddite here. Has anyone with decent (5+ years) experience built a non-trivial new feature in a production codebase quicker by letting AI write it?
Agents are great at familiarizing me with a new codebase. They're great at debugging because even when they're wrong, they get me thinking about the problem differently so I ultimately get the right solution quicker. I love using it like a super-powered search tool and writing single functions or SQL queries about the size of a unit test. However, reviewing a junior's code ALWAYS takes more time than writing it myself, and I feel like AI quality is typically at the junior level. When it comes to authorship, either I'm prompting it wrong, or the emperor just isn't wearing clothes. How can I become a believer?
This article was a bit confusing for me. It starts off by describing what "doing it wrong" looks like (okay). It then goes on to talk about Agents. Perhaps it's just that my human brain needs a firmware update, but I was expecting the "what doing it wrong looks like" section to be followed by a "what doing it right looks like" section. Instead, the next paragraph just begins with "Agents".
Sure, one could surmise that perhaps "doing it right" means "using Agents", but that's not even how the article reads:
> "To make AI development work for you, you’ll need to provide your AI assistant with two things: the proper context and specific instructions (prompts) on how to behave under certain circumstances."
This, to me, doesn't necessitate the usage of agents, so to then enter a section of agents seems to be skipping over a potentially-implied logical connection between the problem in the "doing it wrong" section and how that is solved in the "Agents" section.
Copying code snippets into web UIs and testing manually is slow and clunky, but Agents are essentially just automations around these same core actions. I feel this article could've made a stronger point by getting at the core of what it means to do it wrong.
• Is "doing it wrong" indicated by the time wasted by not using an agentic mechanism vs manual manipulation?
• Is "doing it wrong" indicated by manually switching between tools instead of using MCP to automate tool delegation?
Having written several non-trivial agents myself using Gemini and OpenAI's APIs, the main difference between handing off a task to an agent and manually copy/pasting into chat UIs is efficiency — I usually first do a task manually using chat UIs, but once I have a pattern established, or have identified a set of tools to validate responses, I can then "agentify" it if it's something I need to do repeatedly. But the quality of both approaches is still dependent on the same core principles: adequate context (no more nor less than what keeps the LLM's attention on the task at hand) and adequate instructions for the task (often with a handful of examples). In this regard, I agree with the author, as correct context + instructions are the key ingredients to a useful response. The agentic element is an efficiency layer on top of those key ingredients which frees up the dev from having to manually orchestrate, and potentially avoids human error (and potentially introduces LLM error).
6 comments
[ 2.9 ms ] story [ 24.3 ms ] threadI would've thought that following the initial argument and the progression to the latest trend we would've ended at use agents and write specs and these several currently popular MCPs.
I guess my rant is it to arrive at the point that no one knows what the "correct" way to use them is yet. A hammer has many uses.
Agents are great at familiarizing me with a new codebase. They're great at debugging because even when they're wrong, they get me thinking about the problem differently so I ultimately get the right solution quicker. I love using it like a super-powered search tool and writing single functions or SQL queries about the size of a unit test. However, reviewing a junior's code ALWAYS takes more time than writing it myself, and I feel like AI quality is typically at the junior level. When it comes to authorship, either I'm prompting it wrong, or the emperor just isn't wearing clothes. How can I become a believer?
Sure, one could surmise that perhaps "doing it right" means "using Agents", but that's not even how the article reads:
> "To make AI development work for you, you’ll need to provide your AI assistant with two things: the proper context and specific instructions (prompts) on how to behave under certain circumstances."
This, to me, doesn't necessitate the usage of agents, so to then enter a section of agents seems to be skipping over a potentially-implied logical connection between the problem in the "doing it wrong" section and how that is solved in the "Agents" section.
Copying code snippets into web UIs and testing manually is slow and clunky, but Agents are essentially just automations around these same core actions. I feel this article could've made a stronger point by getting at the core of what it means to do it wrong.
• Is "doing it wrong" indicated by the time wasted by not using an agentic mechanism vs manual manipulation?
• Is "doing it wrong" indicated by manually switching between tools instead of using MCP to automate tool delegation?
Having written several non-trivial agents myself using Gemini and OpenAI's APIs, the main difference between handing off a task to an agent and manually copy/pasting into chat UIs is efficiency — I usually first do a task manually using chat UIs, but once I have a pattern established, or have identified a set of tools to validate responses, I can then "agentify" it if it's something I need to do repeatedly. But the quality of both approaches is still dependent on the same core principles: adequate context (no more nor less than what keeps the LLM's attention on the task at hand) and adequate instructions for the task (often with a handful of examples). In this regard, I agree with the author, as correct context + instructions are the key ingredients to a useful response. The agentic element is an efficiency layer on top of those key ingredients which frees up the dev from having to manually orchestrate, and potentially avoids human error (and potentially introduces LLM error).
Am I missing something here?
By the end of the current generation’s careers in a few decades surely it will be able to do everything.
I can't believe how far I got on that article before it finally dawned on me that it's just some AI slop. (Well, that's the charitable explanation.)