We’re currently in the “trough of disillusionment” phase of the AI agent hype cycle, and I think that’s actually great news.
A year ago, every AI company was pitching fully autonomous agents that would replace entire departments. “Set a goal and walk away — the agent handles everything.” Demo videos showed agents smoothly navigating complex workflows, making decisions, and producing perfect outputs. VCs poured billions into agent startups.
Today, most of those demos haven’t translated to production systems. The companies that bought into “fully autonomous everything” are quietly scaling back to “AI-assisted workflows.” The agents that were supposed to handle complex multi-step tasks reliably turn out to handle them unreliably about 30% of the time — which, in production, means they’re unreliable 100% of the time.
This isn’t failure. This is normal technology adoption. And understanding where we actually are on the curve tells you what to bet on now.
What Actually Works in 2026
AI-assisted workflows: very mature. Human does the thinking, AI handles the tedious parts. Writing drafts, summarizing documents, analyzing data, generating code suggestions. This is the “electricity” phase — it’s so embedded in daily work that we’re already forgetting what it was like without it.
Scheduled automation: reliable. AI agents running on schedules — morning briefings, daily reports, weekly summaries, monitoring checks. These work because they’re predictable: same task, same time, same format. The AI doesn’t need to make complex decisions; it needs to execute well-defined tasks consistently.
Simple reactive agents: solid. Agents that respond to specific triggers with specific actions. “When someone asks about X in Slack, provide answer Y.” “When a new PR is opened, generate a review summary.” Single-step responses to clear triggers. Reliable enough for production.
Complex autonomous agents: not there yet. Multi-step workflows where the agent makes decisions about what to do next based on intermediate results. “Research this market, identify the best opportunity, create a strategy, and build a presentation.” Each step is fine individually. The orchestration — deciding what comes next based on what happened in the previous step — is where things fall apart.
The failure mode isn’t dramatic. The agent doesn’t crash or refuse. It just makes subtly wrong decisions about what to do next. It decides a tangent is worth exploring when it isn’t. It misinterprets an intermediate result and goes down the wrong path. It produces plausible-looking output that’s based on flawed reasoning. These failures are harder to catch than crashes, which makes them more dangerous.
What’s Coming in the Next 12-18 Months
Better tool use. Models are getting significantly better at using tools — making API calls, querying databases, manipulating files. This is the foundation for more reliable autonomous agents. When the tool use layer is rock-solid, the orchestration layer can be thinner and simpler.
Smaller, specialized agents. Instead of one mega-agent that handles everything, we’ll see collections of small, specialized agents that each do one thing really well. A code review agent. An invoice processing agent. A customer support triage agent. Each one is narrow enough to be reliable.
Better evaluation and testing. We’re getting better at measuring agent performance systematically. Instead of “it seemed to work in a demo,” we’ll have benchmarks, test suites, and confidence scores that tell you how reliable an agent actually is for your specific use case.
Human-in-the-loop as a feature, not a limitation. The narrative is shifting from “the agent should be fully autonomous” to “the agent should be autonomous for routine cases and escalate to humans for edge cases.” This is more realistic and produces better outcomes.
What This Means for You
If you’re buying AI tools: buy the boring ones. AI-assisted writing, AI-powered search, AI-generated summaries — these deliver value today, reliably. Skip the “autonomous AI that replaces your [role]” pitch for another year.
If you’re building AI tools: build for the current reality, not the demo. A tool that reliably handles 80% of a workflow and gracefully hands off the remaining 20% to a human is more valuable than a tool that promises 100% autonomy but delivers 70% accuracy.
If you’re investing in AI: look for companies solving specific, well-defined problems rather than building general-purpose autonomous agents. The specific-problem companies will generate revenue now. The general-purpose agent companies are mostly burning cash waiting for the technology to catch up to their vision.
The Prediction I’m Most Confident About
By the end of 2027, the most successful AI agent companies won’t be the ones that achieved full autonomy. They’ll be the ones that found the right balance between automation and human oversight for specific, high-value workflows.
The fully autonomous agent dream isn’t dead. It’s just further away than the hype suggested, and the path there goes through “really good human-AI collaboration” rather than “replace humans entirely.”
And honestly? The collaboration path produces better outcomes anyway. A human with AI tools beats a fully autonomous AI agent on any task that involves judgment, creativity, or navigating ambiguity. Which is most tasks that actually matter.
🕒 Last updated: · Originally published: December 11, 2025