The demos look good. An AI agent browses the web, drafts an email, confirms a hotel booking - all without you touching the keyboard. The companies building these systems have staked their near-term revenue on the idea that agents are ready to handle real work. The reality, at least right now, is narrower and more frustrating than the pitch.
The failure modes aren’t dramatic. There’s no Skynet moment. What actually happens is that an agent successfully completes four out of five steps in a task, then stalls on a dropdown menu that behaves slightly differently than expected - a shadow DOM element, a CAPTCHA, a session timeout. It doesn’t fail loudly. It often just stops, or worse, silently does the wrong thing and reports success.
This is the problem with how agents get evaluated. The popular benchmarks - WebArena, GAIA, and others in that space - measure task completion on controlled environments. Real software is not a controlled environment. Forms break. APIs return unexpected errors. A customer service portal might display a modal that wasn’t there last week. Agents trained and tested in cleaner settings fall apart when the real world doesn’t cooperate, and there’s no standard way to measure how gracefully they fail.
The Brittleness Nobody Advertises
Computer-use agents from Anthropic and others made a visible splash when they launched, and they genuinely work on a subset of tasks. But the subset is defined by interface predictability. Standard buttons, consistent layouts, well-structured pages - these things help. Legacy enterprise software, government portals, and anything built before 2015 tends to wreck them.
Enterprise vendors selling agentic workflows are aware of this. The quiet workaround is heavy human-in-the-loop design: the agent handles what it can, a human reviews what it can’t. That’s a reasonable approach, but it’s also not what the top-line marketing says.
Short Loops, Not Long Ones
Where agents actually hold up is in short, tightly scoped tasks with predictable inputs. Summarising a document, extracting structured data from an email, generating code from a spec with well-defined output formats. The longer and more contingent the task chain, the faster reliability degrades. Each additional step is another surface for compounding error.
The underlying models keep improving, and some of the interface brittleness will be solved by better tool design rather than smarter agents. But the version of agentic AI being sold in 2026 - autonomous, reliable, end-to-end - is running ahead of what actually ships. The gap isn’t secret. It’s just not what the demos show.