A recent post on the Cursor forum questioning the push for agentic models, when they still struggle to follow simple instructions, has reignited one of the most grounded debates in software development.
Can AI truly be trusted to write or even manage code autonomously?
The frustration behind the post was evident. When a developer asked GPT-5 and Gemini Pro to update a Go function, both models failed to make even basic updates. For many engineers, this reflects the current reality of agentic coding—messier than promised, and far from trustworthy.
Yet, to dismiss the idea as “useless” might be premature. Inan exclusive chat with AIM, Namanyay Goel, founder of Giga AI, and Brijesh Patel, CTO of SNDK Corp, argued that agentic coding isn’t a pipe dream but rather a tool that must be appropriately scoped and supervised.
Where Agentic Coding Actually Works
Goel describes agentic coding as a “proposal system” rather than an autonomous one. He explained that its strengths lie in “the boring stuff” such as scaffolding, authentication, CRUD endpoints and database migrations. These are repetitive areas where the patterns are well understood and errors are low-impact.
“When the problem has been solved a million times before, the agent nails it,” Goel noted.
Prototyping is another sweet spot. When a company needs to test several ideas quickly, agents can scaffold multiple prototypes in hours. “Two will probably fail, and that’s fine,” he added.
“Speed from AI, judgment from humans—that’s the balance,” he said.
Goel also pointed out that small, isolated services such as payment webhooks and email notifications are ideal for agents. These are clear-boundary tasks, so even if the AI fails, it does not jeopardise the integrity of the wider codebase.
Developers on Hacker News have echoed similar views. One user noted that while LLMs occasionally make errors, they “more often than not follow instructions” when given well-structured prompts.
Another compared the experience to “working with a smart intern who needs hand-holding”, adding that once guided effectively, “it can fly through boilerplate tasks”.
A third developer expanded on the analogy, observing that, like a smart and capable intern lacking real-world experience, guiding an AI agent to perform technical tasks correctly demands significant human effort and time. However, unlike human interns, AI agents operate at a much faster pace.
The Reality Check: Where It Fails
The limits of agentic coding become evident the moment decisions require architectural judgement. Goel cautions that agents “write for the happy path” and fail to consider “what happens when your database grows 100x.”
They may produce code that compiles, but not code that scales or anticipates real-world complexities such as caching, concurrency or load handling.
Echoing this caution, Patel said, “AI not being entirely trusted for coding, combined with stressed code reviews, raises important concerns.”
“However, agentic coding, where AI operates in the background to assist developers without fully taking control, is a promising middle ground.”
His team uses AI agents to handle generic coding, routine code checks and documentation, freeing human engineers to concentrate on architecture and business logic.
Several developers on Hacker News agreed that if they must spend the same amount of time explaining a problem and validating the AI’s output, “it’s not a time saver, it’s a time sink.”
Another added that while agents are good at rewriting or translating simple code, “the moment I want something slightly off the beaten path, it struggles and makes mistakes.”
Also Read: Agentic DevOps Helps Developers ‘Tear Through Crushing Technical Debt’
The Human-in-the-Loop Equation
Both leaders agree that agentic systems fail when humans surrender oversight. “Human-in-the-loop will always be critical to maintain control, ensure quality, and interpret complex business logic,” Patel insisted.
While agentic coding can complicate reviews, as reviewers must validate both the AI’s code and its decision-making process, he believes the net effect improves efficiency in the long run. “Thoughtful design can reduce repetitive review tasks and highlight high-risk areas that need attention,” he explained.
Goel’s checklist for this balance is pragmatic. Developers, he says, should treat agents as proposal systems rather than guarantees, reviewing all generated code before it goes live. They should assign agent work to separate branches, enforce test coverage and focus code reviews on critical aspects such as error handling, validation and concurrency. The goal is not blind delegation but structured collaboration between AI and engineers.
Also Read: Why Programmers Are Using Problematic AI Code?
Beyond the Hype, Towards Discipline
For developers still sceptical, both experts suggest reframing the narrative. The goal is not full autonomy but semi-autonomous assistance. Patel foresees a gradual evolution in the coming years. He stressed that while AI-powered tools are improving, their broad adoption hinges on building trust, ensuring transparency and seamless integration into workflows.
In Goel’s words, “Agentic coding doesn’t eliminate the need to understand what you’re building. It’s a multiplier for people who know what good code looks like.”
The verdict, then, lies somewhere between hype and dismissal. Agentic coding proves useful when it is carefully scoped, supervised and grounded in human judgement. The real trouble begins only when teams forget that automation without understanding is just noise, no matter how fast it runs.
The post Is Agentic Coding Just a Smart Intern Pretending to Know More? appeared first on Analytics India Magazine.
