I’ve been watching the AI agent space closely for about two years. Long enough to have gotten genuinely excited, then genuinely skeptical, and now something in between that I think is closer to accurate.
The pitch for AI agents goes something like this: instead of answering a question, an AI takes a sequence of actions to accomplish a goal. It searches the web, writes code, runs it, checks the output, fixes errors, calls APIs, sends emails. It doesn’t wait to be asked — it works through a task the way a person would.
That pitch is real. Agents can do that. I’ve seen it work. The question is when it works, how reliably, and whether the reliability is good enough to stake anything on.
Where Agents Actually Deliver
The clearest wins are in bounded, well-defined tasks with fast feedback loops. Code execution is the best example. An agent that writes a Python script, runs it, sees an error, reads the traceback, and fixes the bug — that loop is tight, the feedback is unambiguous, and the agent can iterate until something works. I’ve watched Claude Code do this for 20 minutes on a build problem I would have spent an hour on. The agent has infinite patience for mechanical iteration.
Data extraction and transformation is another strong category. Give an agent a pile of PDFs, a schema, and a destination, and it will grind through it. Not always perfectly, but consistently enough to be worth running unattended on tasks you’d otherwise hate.
Web research with specific, verifiable outputs also holds up. “Find the last three earnings call dates for these five companies and put them in a CSV” is the kind of request where an agent can go do actual work rather than produce plausible-sounding text.
The pattern: agents work when the task has clear stopping conditions and the agent can tell whether it’s succeeded.
Where They Don’t
Long chains of reasoning that depend on getting intermediate steps right. Errors compound. An agent that misunderstands the first step in a ten-step process doesn’t usually self-correct — it confidently executes the wrong plan. The longer the chain, the more places things can quietly go sideways before you see an obviously wrong output.
Anything requiring judgment about what the user actually wants, as opposed to what they literally said. Agents are terrible at “wait, this doesn’t seem right, let me check before continuing.” They continue. That’s what they do. It’s on you to build in checkpoints.
Tasks where “wrong but plausible” is worse than “not done.” A lot of business workflows fall into this category. An agent that drafts a contract, sends an email, or modifies production data has a much higher cost of failure than an agent that runs a unit test. The irreversibility matters enormously and agents don’t naturally weight it.
The Reliability Problem Is Real
Enterprise adoption is moving slower than the demos suggest, and I think the reliability gap is the real reason. Not because the demos are fake, they’re usually real. But a tool that works 85% of the time on a task you run hundreds of times a day is a tool that fails constantly.
Benchmark scores don’t translate cleanly to production reliability. An agent that scores well on SWE-bench (a software engineering benchmark) is being tested on problems that were written specifically to have correct answers verifiable by running tests. Production software development has a lot more ambiguity than that. “Make this perform better” is not a benchmark problem.
The research on agentic task success rates shows a consistent pattern: single-step tasks go well, multi-step tasks degrade. Longer task horizons mean more opportunities for the model to hallucinate, misread its own previous output, or get stuck in a loop. The frontier models are getting better at this — meaningfully better — but they haven’t closed the gap enough for unsupervised long-horizon work in most domains.
The Orchestration Tax
Building a reliable agent isn’t just prompting an LLM and hoping for the best. You’re building infrastructure. Retry logic, state management, error handling, tool definitions, context window management, cost monitoring, output validation. The agent frameworks (LangChain, LlamaIndex, others) help, but they also abstract away things you need to understand when something breaks.
I’ve talked to developers who spent three months building an agent pipeline for a workflow that a well-engineered traditional script would have handled in two weeks. The agent version is more impressive to demo. It’s also harder to debug, more expensive to run, and fails in stranger ways. For some use cases, that tradeoff is worth it. For a lot of them, it isn’t.
The enthusiasm around agents sometimes skips this part. You’re not just calling an API — you’re taking on operational complexity, and that complexity scales with the ambition of what you’re trying to automate.
What I Think Is Worth Building Now
Human-in-the-loop agents, with the human inserted at the genuinely risky decision points rather than at every step. The goal isn’t full automation — it’s handling the tedious 80% automatically while surfacing the 20% that needs judgment. That’s a lot more achievable than full autonomy and often more valuable.
Agents with narrow, well-scoped tool access. An agent that can read files, write files, run code, and call one specific internal API is safer and more predictable than one with a hundred tools it might invoke inappropriately. Fewer tools, clearer task scope, better results.
Workflow automation where failure is detectable. If your agent is doing something reversible, or something where errors surface quickly and clearly, you can let it run with confidence. If it’s making decisions that are hard to audit after the fact, add the checkpoint.
I’m genuinely interested in what agents can do. The best uses I’ve seen aren’t replacing human workflows wholesale — they’re compressing the mechanical parts of workflows that were already well-understood. That’s real value. It’s just not the “agents will do everything” narrative that tends to dominate the announcements.
Late 2026 is a good time to be skeptical and curious at the same time. The technology is genuinely capable. So is the hype.