The AI agent industry is caught in a cycle of constant reinvention, where new terminology and product names create a false sense of obsolescence rather than genuine progress. Peter Steinberger, creator of OpenClaw, recently highlighted this problem by questioning whether the industry has moved from discussing "loops" to "graphs," pointing to a deeper issue: people are following architectural trends without clearly understanding what problems they actually solve.

The confusion stems partly from how AI agent products are named and marketed. ChatGPT, Claude, OpenClaw, Atlas, and dozens of other systems can all perform similar tasks, yet they are presented as a linear progression from less advanced to more advanced. This creates an uncomfortable sense of obligation among users and developers, who feel pressured to adopt the latest tool before their current workflow becomes obsolete. The architecture becomes the goal, rather than solving a specific problem.

What Actually Makes Up an AI Agent?

To cut through the noise, Steinberger proposes breaking down agentic systems into their actual components rather than treating them as monolithic products. Understanding these layers helps explain why two agents with different names might be more similar than they appear, and why a newer product isn't automatically better for your use case.

• Language Model: The reasoning engine that interprets requests and produces language, code, or tool calls. GPT and Claude model families belong here, but the product name doesn't always reflect which model is inside. An agent calling Claude through an API is not the same as using Claude Code's full harness.
• Context and Memory: Context is what the model can see during a single run, including prompts, conversation history, files, and search results. Memory is information the system preserves for later use, such as preferences or project facts. These are separate layers; context is temporary input, while memory persists across sessions.
• Tools and Skills: Tools let the model reach beyond its context, such as web search, file access, or connectors to Gmail and Slack. Skills package instructions and workflows for repeatable tasks. The distinction matters: instructions about reading Gmail are not the same as Gmail credentials.
• Harness and Execution Environment: The harness is the runtime machinery that assembles prompts, offers tools, executes commands, and decides when a job is finished. The execution environment determines where work happens, whether in the cloud, a virtual machine, a local folder, or a Git repository. This layer often matters more than a small difference between models.
• Permissions and Interface: Permissions determine what the agent can actually do with the resources it can reach. The interface is simply the doorway, whether chat, CLI, IDE extension, or desktop app. Two products that look identical in chat may lead to completely different tools and permissions underneath.

How to Evaluate an AI Agent for Your Actual Needs?

Rather than chasing the latest product name, Steinberger suggests a practical approach: define the job you need done, then examine whether each candidate system can actually do it. He tested this method by asking multiple agents to prepare a daily brief from current news and several email accounts, keeping the result fixed while varying the tools.

• Start with the outcome: Define exactly what you need the agent to produce. A daily brief, a code review, a research summary, or something else. This keeps you focused on results rather than terminology.
• Check the execution environment: Can the agent access the files, accounts, and services you need? Does it run locally on your computer, in the cloud, or in a browser? This often matters more than which language model is inside.
• Verify permissions and safety: What is the agent actually allowed to do? Can it read your inbox without sending mail? Can it write to one folder but not inspect the rest of your disk? Permissions cut across the harness, tools, and environment.
• Understand the harness, not just the model: Knowing that an agent uses GPT-4 tells you about its reasoning ability, not whether it can save files where you want them or integrate with your existing workflows. The harness is what makes it an agent.
• Ignore the product name as a guide: Claude Code, Claude, and an agent that calls Claude through an API are different architectural claims. The name alone doesn't tell you which components are inside.

Why Does This Matter Right Now?

The AI agent landscape is moving fast, and products are changing substantially within months. Some are scheduled to disappear entirely. In this environment, understanding the underlying components of an agent is more valuable than memorizing product names. When you know what a harness does, what tools can access, and how permissions work, you can evaluate new products on their actual merits rather than their marketing.

Steinberger's framework also addresses a real frustration: the sense that you are always behind. Prompt engineering gave way to context engineering, which gave way to harness engineering, and now loops are supposedly being replaced by graphs. Each term describes something real, but without clear definitions, they feel like a sequence of replacements where last week's knowledge is already obsolete. By separating the components, you can see that many of these terms describe different layers of the same system, not a linear progression.

"Are we still talking loops or did we shift to graphs yet?" Steinberger asked on X, capturing the discomfort with how the industry presents each new concept as a replacement for the last one.

Peter Steinberger, Creator of OpenClaw

The practical implication is clear: before adopting a new agent or moving an existing workflow to a new platform, ask what specific problem it solves for you. If your current setup handles the job, the fact that a newer product exists is not a reason to switch. The architecture should serve the work, not the other way around. In an industry that moves as quickly as AI, that kind of clarity is increasingly valuable.