Agentic AI Coding: A Dream Come True, or a Nightmare?

It's June 2026, and agentic AI tools for programming have become the standard. According to an analysis by Joe Bertolami of VentureBeat, CTO of Clifton AI, agentic systems today handle most routine code generation. GitHub Copilot, Anthropic's Claude Code, and Anysphere's Cursor have become an integral part of developer workflows — even in Czech companies.

However, the result is paradoxical. While code is increasing at a record pace, product quality is not growing at the same rate. "If we're delivering code faster than ever before, why aren't our products improving at the same speed?" — this is a question Bertolami hears from business leaders more and more frequently.

The answer is surprisingly simple: writing code was never what hindered development. What was — and remains — the bottleneck is the correct definition of requirements, integration with complex systems, and software maintenance in real-world conditions. And when agentic AI starts churning out hundreds of lines of code daily, all these problems only deepen.

The New Bottleneck: Human Review

With the growing volume of AI-generated code, human review (code review) is becoming the main impediment to development. Developers suddenly can't keep up with checking everything their AI assistants have produced. And what's worse — they lose the context needed to detect errors made by the agents.

"Agents compress the time needed for execution. But they don't compress ambiguity, accountability, or operational complexity," Bertolami explains. In other words: AI writes code in minutes, but verifying that it does what it's supposed to takes hours.

This problem has already escalated from theoretical debate to real numbers. According to an Axios investigation from late May 2026, one unnamed corporate client spent half a billion dollars in a single month on Claude licenses after failing to set limits on their use. Microsoft canceled most Claude Code licenses for its employees — partly due to costs. And Uber, according to Business Insider, admitted that it had exhausted its AI budget for 2026 by April.

Where It Breaks: A Playbook for Companies

In his analysis, Bertolami proposes a three-phase plan for managing the transition to agentic development without the company going bankrupt or losing control over its codebase:

1. Financial and Security Governance

Never let an agent inherit the full permissions of its human operator. A human developer has broad access because they bear responsibility. An agent does not. Bertolami recommends strictly separating read and write access and requiring human approval for destructive or production actions. At the same time, it is necessary to set quotas and limits for AI consumption — otherwise, the "half a billion a month" scenario looms.

2. Technical Strategy: Multi-model and Measuring the Right Metrics

No single model excels at everything. Organizations should use multiple models from different vendors and precisely map which model is suitable for which type of task. Standardizing on a single vendor creates a critical "single point of failure."

At the same time, it's time to stop measuring developer productivity by the number of lines of code, deployments, or pull requests. With AI, these metrics are misleading. Instead, Bertolami recommends tracking metrics tied to business outcomes: feature adoption, retention, change failure rate, or "code survival over time."

3. Talent and Organization: From Syntax Scribes to System Architects

This is perhaps the most crucial point for Czech developers. Companies must retrain their people from "syntax scribes" to "system thinkers." When agents handle most code generation, human added value shifts to architecture, integrations, and oversight of agents.

Bertolami warns against hasty layoffs: "If you haven't integrated agentic workflows, measured augmented output in production, and re-engineered your roadmap around faster execution, you don't actually know if your needs and capacities align."

Czech Context: What Does This Mean for Local Companies?

In the Czech Republic, AI tools for development are widely available. GitHub Copilot costs 100 dollars per month in its new Copilot Max tier starting October 2025. Anthropic's Claude Code is billed based on token consumption. Cursor offers a free tier with limited requests and paid plans starting from 20 dollars per month.

For smaller Czech companies and startups, this applies doubly: without governance set up from the start, the AI bill can quickly exceed the budget. And without a clear code review strategy, there's a risk that three developers with AI agents will produce so much code that no one will be able to review it.

At the same time, there's an opportunity. Czech companies that can combine European quality and meticulousness with the speed of AI agents can gain a competitive advantage over their American counterparts, who often rely on quantity over quality. According to Axios, the reality of AI in 2026 is that "the only place where AI truly works is programming" — and even there, with reservations.

Box Created 13 New Positions. Is This a Harbinger?

An interesting precedent was set by Box, which, according to The New York Times, created 13 entirely new types of job positions directly due to the advent of AI. These are not roles replacing developers — they are roles that oversee what AI produces. AI auditor, AI quality lead, agent workflow designer — professions that didn't exist two years ago.

For the Czech market, this means a clear signal: whoever learns today to manage, control, and architecturally oversee AI agents will be indispensable tomorrow.