In the wild west of AI marketing, the term 'agent' has become a catch-all for everything from simple chatbots to fully autonomous systems. Ryan Kowalski’s updated October 2026 analysis cuts through the noise by focusing on what is actually running in production versus what is merely a demo. The core insight is simple but critical: a system’s value isn't defined by its label, but by its bounded scope, permission levels, and the clarity of its escalation paths. If you can't define who reviews the output, you're not deploying an agent; you're deploying a liability.

Production Deployments vs. Vendor Claims

The article highlights three distinct categories of deployed systems: Klarna’s customer-service assistant, coding tools like Devin, and John Deere’s See & Spray machinery. These aren't interchangeable. Klarna reported handling 2.3 million conversations in its first month in February 2024, equating that workload to 700 full-time agents. However, this was a company claim, not an independent audit. By May 2025, Bloomberg reported Klarna was reintroducing more human support options, proving that high volume does not equal high quality or full autonomy. The lesson for engineers is to treat vendor workload figures as marketing metrics until they are backed by resolution rates and repeat-contact data.

The Coding Agent Paradox

For those building dev tools, the evidence is even murkier. While benchmarks scream 'revolution,' real-world data tells a different story. METR’s July 2025 randomized study found that experienced developers took 19% longer to complete tasks when using early-2025 AI coding tools, despite expecting a speedup. This study, involving 16 developers and 246 tasks, exposes a dangerous gap between benchmark performance and workplace productivity. A rising score on a static dataset does not account for the cognitive load of reviewing, integrating, and fixing AI-generated patches. As of October 2026, the consensus is clear: coding agents are useful for drafting, but they are not yet autonomous engineers.

Market Reality: Agent Washing and Cancellation

The hype cycle is already showing cracks. Gartner projected in June 2025 that over 40% of agentic AI projects would be canceled by the end of 2027 due to unclear value and rising costs. Perhaps more damning, Gartner estimated that of the thousands of vendors claiming to sell 'agents,' only around 130 were doing something substantive, labeling the rest as 'agent washing.' McKinsey’s November 2025 survey further contextualizes this, noting that while 62% of organizations reported some form of agent use (including experimentation), only a tiny fraction met the definition of an 'AI high-performer' with significant EBIT impact. Adoption is widespread; success is rare.

Key Takeaways

  • Autonomy is a spectrum: Check permissions, stopping conditions, and human-in-the-loop requirements before trusting any 'agent'.
  • Volume is not value: Klarna’s 2.3 million conversations included a pivot back to human support, proving scale doesn't solve quality.
  • Benchmarks lie: METR data shows experienced devs slowed down by 19% with early AI tools, highlighting the review overhead.
  • Agent washing is rampant: Gartner estimates only ~130 vendors out of thousands are offering substantive agentic solutions.
  • Start small: Successful deployments are bounded tasks like email triage or invoice extraction, not entire department replacements.

The Bottom Line

Stop looking for a fully autonomous agent to replace a department. Instead, look for bounded workflows with clear escalation paths and human approval gates. The agents that actually work today are boring, supervised, and specific—not the flashy, general-purpose robots sold in press releases.