Erik Johannes has spent two years deep-diving into AI agents as part of publicly funded academic research, and his conclusion cuts against the relentless hype machine: the actual usefulness of these tools is more limited than most people think. Writing on his personal site in April 2026 (before catching fire on Hacker News), Johannes makes a case that's equal parts philosophical critique and hands-on technical assessment. His core thesis lands hard: AI agents don't give us better intelligence—they give us increased speed. In a world obsessed with throughput, that sounds like a win. But Johannes argues we're forgetting we live in bodies with biological limits.

The Productivity Paradox

Johannes echoes Ed Zitron's skepticism about whether AI can do anything beyond making "some engineers do some stuff faster." The missing link, he argues, is that simple speedups don't automatically translate to value. You can't measure value in lines of code or development velocity—those are outputs, not outcomes. This tension between American-style productivity obsession and more holistic European perspectives on what "value" actually means runs throughout the piece. Johannes doesn't pull punches: "There is undoubtedly a huge difference in how we view what 'value' actually means."

Coding Agents: The One Place They Actually Work

Here's where things get interesting for builders. Johannes tested GitHub Copilot, OpenAI's Codex, Claude Code, and Goose extensively—and admits programming is "by far the most promising use case." But it wasn't always that way. In fall 2025, he found coding agents "very unwieldy, producing unnecessary amounts of code, quickly making whole projects unmanageable." Inline autocomplete felt preferable to full agentic autonomy. Now? He can build smaller prototypes while maintaining oversight. The key: strict guardrails. Johannes developed his own instruction set for keeping agents focused—maximum 1-2 files per request, always state your plan before writing code, wait for approval on changes over 100 lines.

Research Agents Still Miss the Point

Johannes tested an AI agent's ability to produce a research paper from literature review to final PDF. The result? Coherent but not interesting or relevant. "The interpretation of the results is not so simple to outsource," he notes, and even if an LLM can generate consistent discussion, there's still the question of whether results are actually interpreted if no human deemed them useful first. This gets at something deeper: as we push more automation into research, we risk reducing human agency rather than achieving the self-empowerment AI vendors promise.

The Ethical Bill Is Coming Due

Beyond practical limitations, Johannes tackles the moral dimension that too many in tech want to sweep under the rug. Copyright infringement and worker exploitation in LLM training "have real negative consequences for real people," he writes. Agentic systems also massively accelerate energy consumption through token generation at scale. He cites Moltbook (a social network for AI agents) and ClawXiv (research papers written by AI agents) as examples where benefit versus cost is "exceptionally skewed." The price is too high, but hidden from most users.

Why Johannes Still Uses Them Anyway

The uncomfortable truth: Johannes would prefer not to use LLMs at all due to ethical concerns. But over the last few months, that's changed for programming specifically. "A new normal has arrived for most jobs involving software development." He acknowledges this is a compromise driven by economic reality, even while believing slower, more thoughtful development would be more sustainable. "The exploitative practices and inhumane working conditions of mineral extraction lay the foundation not only for data centers running gigantic AI models, but also the laptop that I type on."

Key Takeaways

  • Coding agents have matured significantly since late 2025, but require strict guardrails to remain useful rather than destructive
  • Speed is AI's main contribution, not intelligence—and that's a crucial distinction for evaluating real value
  • Research automation raises questions about human agency that go beyond simple efficiency metrics
  • Ethical costs (copyright, labor, energy) are systematically hidden from end users

The Bottom Line

Johannes isn't anti-AI—he's paid to study it. But his two-year assessment reveals a tech industry desperately selling speed as transformation while the actual intelligence gap remains wide. For developers: use coding agents with guardrails and maintain oversight, or they'll own your codebase before you realize it.