When Agents Do More Than Code: Who Is Actually Paying?

The most valuable information in this article is that it admits last year's hype about "AI coding landing fast" needs correction now. But the direction of the correction might be more pessimistic than they imagine.

At this time last year, everyone said AI was landing fastest in the coding field, and Copilot's efficiency gains were verified. Now Agents have arrived, upgrading from assistance to autonomous mode, starting to enter enterprise workflows. Sounds great, but look closely at those cases in the news—do you notice a commonality? They all stay at the "demo" and "pilot" stage; none have truly achieved large-scale enterprise deployment.

An AI company CTO said in an interview: "We've already used Agents internally to handle 30% of code review tasks, with significant efficiency improvements."

But 30% of code reviews is still an ocean away from "taking over enterprise workflows." My own company tried something similar, only to find that while Agents could indeed handle simple repetitive CRs (code reviews) in internal tests, they started hallucinating as soon as business logic judgment or understanding historical legacy code context was involved. Let alone code changes requiring cross-department collaboration or compliance reviews—the Agent couldn't even understand comments regarding security policies.

This exposes two core problems behind the scenes.

First, the Agent's "autonomous mode" is actually a false proposition. So-called autonomy just automates steps previously triggered manually, but decision boundaries still rely on preset rules. Once encountering exceptions outside the rules, it either gets stuck or gives wrong suggestions. I personally witnessed an Agent changing database connection pool configurations to production parameters during refactoring, nearly causing an accident. Post-mortem analysis revealed the Agent fundamentally didn't understand the basic concept of "environment isolation"; it just executed the instruction "optimize connection pool parameters" literally.

Second, enterprise workflows aren't just piles of code. The news says Agents are "entering enterprise office scenarios," which reminds me of last year's hype about AI writing PPTs. In reality, any automation involving organizational processes, permission management, and legacy system compatibility is far more complex than code generation. An Agent might be great at writing a Python script, but ask it to understand an enterprise's procurement approval process, and it can't even answer basic questions like "who has authority to approve."

[!tip] What's truly alarming is that capital is packaging Agents as a "master key," but enterprises' actual need is a "locksmith." Every organization has different locks; this Agent key looks universal but can actually only open a few standard locks.

What worries me more is that this hype is misleading startup directions. A team I know was doing vertical-domain AI tools last year but suddenly pivoted to building an "Agent platform" this year, claiming to "enable zero-code workflow automation for enterprises." When I asked them about specific landing scenarios, they couldn't answer, just saying "build the platform first, the ecosystem will form naturally." Isn't this just a rerun of the low-code platform bubble back then? Back then it was also "everyone is a developer," and what happened? Most low-code projects died at the POC (proof of concept) stage.

Returning to Agents themselves, I believe their most valuable short-term application scenario is precisely what the news didn't emphasize: upgrading code assistance tools, not replacing enterprise workflows. Having Agents help developers write unit tests, generate documentation, and do simple code completion—these are real needs with clear ROI. But once crossing this boundary into customer service, sales, finance, etc., Agent reliability, explainability, and compliance become fatal shortcomings.

The news says "Agents are taking over more than just your code." That statement itself isn't wrong. But the word "takeover" is too misleading. It implies a replacement relationship, whereas what enterprises truly need is "collaboration"—humans and AI completing work together, not AI replacing humans. Sadly, current capital narratives emphasize "replacement" because "collaboration" doesn't sound sexy enough to raise money.

My action advice is simple: If you're an enterprise CTO, start by using Agents for development assistance tools, like code review, document generation, and unit testing. These scenarios have low risk and clear benefits. Then, spend at least 6 months observing whether Agents are truly reliable in more complex workflows. Don't get carried away by the four characters "autonomous mode," and don't jump onto an Agent platform you don't understand just to chase trends. Bubbles eventually burst, but enterprises that truly understand Agent capability boundaries will reap the most tangible rewards after the bubble.

Original link: https://www.tmtpost.com/8065779.html