Every vendor on Earth just discovered AI agents. Huawei launched AgentArts. Google Cloud published "AI Agent Trends 2026." Salesforce has "Agentforce." ServiceNow has "AI Agents." Microsoft has "Copilot Agents." The word "agent" has been drained of all meaning, like "cloud" in 2012 or "blockchain" in 2018.
I build AI agent infrastructure. I'm working on Sigil Protocol, an open protocol for agent identity, ownership verification, and reputation. I use Claude Code daily. I orchestrate real autonomous workflows across multiple projects. So when Huawei's AgentArts "cuts agent delivery time by 60%," I have one question: delivery time of what, exactly?
The if-else test
If your "AI agent" just classifies input and routes it to a predetermined workflow, you have built a router with an LLM doing the case matching in a switch statement. Calling it an "agent" is like calling a thermostat "intelligent" because it makes decisions based on temperature.
An actual agent has goals, makes plans, takes actions, observes results, and adjusts. It can fail at step 3 of 7 and figure out an alternative. It can decide it needs information it wasn't given and go get it. It keeps looping and adjusting its plan instead of following a fixed pipeline.
Most enterprise "agents" are pipelines with LLM nodes. They're selling workflows and calling them agents because the word tests better in buyer surveys.
How an agent works
A real agent has a core loop: Observe, Think, Act, repeat. It doesn't know ahead of time how many iterations it needs. Claude Code does this: you give it a task, it reads files, runs commands, hits errors, adjusts, tries again. That's an agent. Your enterprise "agent" that routes support tickets to the right department? That's a two-step pipeline with a marketing budget.
Real agents make judgments about tool selection, deciding which tools to use, when to use them, and whether results are good enough. Workflows are O(n). Agents are O(who knows). That unpredictability is part of what makes them useful.
The memory problem
Most "agents" are stateless functions. They have the memory of a goldfish on a serverless function: born, execute, die, repeat. They keep no context across interactions, so they cannot learn from previous runs or accumulate knowledge. A real agent remembers what it tried, what worked, what failed. It builds a model of its environment over time. These enterprise "agents" can't remember what they did five minutes ago because they weren't designed to. They were designed to process a request and return a response. That gives them the behavior of a function.
Who is responsible for the agent
When agents act autonomously, how do you know who is responsible? Who built this agent? Who deployed it? What has it done before? Can I trust it? Nobody in the "agents eating the world" discourse talks about this. They're too busy writing press releases.
Sigil Protocol solves this with cryptographic identity and verifiable ownership chains. Its reputation system uses actual behavior, since marketing claims are no help when an agent makes a decision that costs money or affects people. Someone needs to be accountable. And "the AI did it" isn't an answer.
Why enterprise vendors love the word "agent"
"Workflow automation" is a 30 dollar per month product and "AI agent" is a 30,000 dollar per year platform. You get the same functionality with a different PowerPoint deck. This is the same thing that was sold as RPA in 2020, "intelligent automation" in 2022, "AI-powered workflows" in 2024, and now "AI agents" in 2026. The architecture barely changed. The pricing tripled.
Where I see real agents
Claude Code and coding assistants that autonomously navigate codebases, run tests, fix bugs, iterate. Research assistants that formulate queries, read papers, synthesize findings, identify gaps, search again. Trading systems that observe markets, form hypotheses, test strategies, adapt. Those are real agents, and building them is hard. Running them costs a lot, with frequent failures along the way. They do not have catchy names or enterprise pricing tiers.
Enterprise vendors want something easy to sell at scale as transformative. Expensive systems that fail often do not fit that pitch. So they took the word "agent" and bolted it onto the same workflow engines they've been selling for a decade.
I expect 2026 to make the word "agent" completely meaningless, despite the "year of agents" pitch. I am happy for people to build agents or workflows. Just be honest about which one you built.