Why AI Agent Recommendations Fail: Not Yet One Category

Opening

At the same moment, two people type different phrases into a search box. One searches “AI agent recommendations,” the other searches “AI agent meaning.”

It’s tempting to assume the person searching for the meaning is a beginner. But in the search-path model I analyzed this time, the link from “meaning” to “recommendations” kept showing up. To be clear, this isn’t a record of any single person actually searching in that order—it’s an estimate of how different search terms connect to one another.

This ordering matters right now. This year, agents have moved past the demo stage and started landing on actual adoption review documents. Those documents need a candidate list and a comparison table—and to build a comparison table, you first have to decide which items belong on the same row.

Reader, I read this ordering as a sign that we still don’t have a clear standard for comparing agent products. Before you can get a recommendation, you first have to decide which tools—with which features and permissions—should even be compared to one another.

In the Search-Path Model, “Meaning” Comes Before “Recommendations”

Let’s start with the numbers.

The average monthly search volume for “AI agent” over the past three months is 29,416. Estimating the 12-month trend by regression gives an annualized rate of +107% (R² 0.515)—interest itself is clearly growing.

What’s interesting comes next. All 15 of the top-ranked model paths I analyzed included “AI agent meaning.” And among the 13 paths where “meaning” and “recommendations” appeared together, meaning came first without exception.

Let me be clear about one thing here. This data isn’t a record of individual clicks—it’s a graph model that collectively estimates how search terms connect to one another. So it would be wrong to read this as “Koreans search for meaning more than for recommendations.” In fact, path frequency itself is slightly higher for “recommendations,” at 367, compared to 347 for “meaning.” What the data lets us read isn’t magnitude—it’s sequence.

The sequence shared across the top 15 paths is worth examining. That said, it’s not a large sample, and it’s not a record of actual individual behavior. This result alone can’t determine the behavior of every buyer or the nature of the market as a whole.

In most product categories, this order is reversed. When we shop for wireless earbuds or laptops, we don’t look up definitions first—we go straight to comparison tables. When a definition search comes first, it means that market hasn’t yet entered the purchase stage.

There’s a counterargument, too. This could be a supply-side effect from the mass production of introductory content targeting “meaning” searches. Even so, a question remains: why is definitional content still being consumed in this market at all? Even with abundant supply, it wouldn’t show up in the path data if there were no demand.

To recommend a tool, you first have to compare tools with the same function and permissions

Recommending a note-taking app isn’t hard. You just compare note-taking apps against each other. The axes of comparison — price, sync, collaboration features — are already shared.

Agents are a different story. Under this one label sit chatbots that only answer questions, automations that run through a fixed sequence, tools that click through browsers on your behalf, and loops that move across multiple systems deciding their own next move. These tools differ so much in what they can be trusted to do and how much autonomy they exercise that ranking them on the same scale is difficult.

Anthropic draws this line relatively clearly. In their framing, a workflow¹ is a system where LLMs and tools are orchestrated through predetermined code paths, while an agent is a system where the LLM directs its own process and tool use. In that same piece, they also admit that different customers define “agent” differently. A vendor issued a definition while simultaneously noting that no consensus exists.

One thing worth flagging: this distinction isn’t an industry-wide legal or academic consensus — it’s a practical line drawn by one major vendor from its own experience. Other companies draw it differently. So when a contract simply says “agent,” that word still defines nothing. When something goes wrong, what matters isn’t the label but the list of permissions and approval procedures attached beneath it. In a market where the definition itself is unsettled, anyone writing documentation should record behaviors, not words.

Evidence that the category is splitting apart also shows up on the technical-standards side. The agent protocol landscape that Google mapped out in March 2026 includes MCP², A2A, UCP, AP2³, A2UI, and AG-UI all at once.

Translate this list from a developer’s perspective into a buyer’s perspective, and it reads like this:

• MCP determines how far this tool can reach into our data. That’s a security review item.
• A2A determines whether our agent talks directly to another company’s agent. That’s a matter of contracts and liability.
• UCP and AP2 handle ordering and payment authority. Finance and internal controls need to look at this.
• A2UI and AG-UI determine how that process appears to human eyes. That’s a matter of audit and user experience.

It’s all called “agent adoption,” yet the review falls to four different departments. That tells you this isn’t one product category — it’s a stack fracturing into multiple layers. Under these conditions, something like “5 Agents Worth Trying” can’t really function as information. A list with no shared axis of comparison isn’t a ranking — it’s just an enumeration.

Same model, different product

There’s a trickier case. Same name, same model inside, but the two end up as different products.

In one configuration, the tool only reads your calendar. It organizes your schedule, shows it to you, and that’s it. In another configuration, the same tool sends out invitations. It builds an attendee list, sends emails, and books a conference room.

On a technical spec sheet, these two aren’t distinguished. The feature list has one identical line: “calendar integration.” But inside an organization, they’re completely different things. If the first one goes wrong, the user just closes the window and it’s over. If the second one goes wrong, it’s already landed in the inboxes of 30 external recipients.

Reversibility is what separates them. Even in Anthropic’s own usage data, irreversible actions made up less than 1% of all tool calls. The share is small, but once executed it’s hard to undo, so it needs a separate confirmation step. That’s a distinct issue from accidents happening in other kinds of calls.

Which is why building a comparison table with product names as row headers fails. You end up needing the same product to sit in two different rows. The row header shouldn’t be the product — it should be the combination of permissions.

You Need to Check How Much You Can Actually Hand Off

What readers really want to know when they check a definition isn’t the dictionary meaning of a word. It’s how much they can safely hand off.

Anthropic’s operational definition, released in April 2026, also explains what the system does on its own. An agent is a system that runs its own loop — planning toward a goal, acting, observing results, and revising — until the task is done or it needs to ask a human. The same document also explains how each tool gets set to always-allow, needs-approval, or blocked.

Here’s where the important part shows up. What actually defines the product isn’t the model name or the marketing copy — it’s the settings screen. So it’s practical to translate the definition along three axes.

Axis	Question to ask	Where to check	Warning sign
Autonomy	Does it choose its own method, or just follow a fixed procedure?	Architecture docs, retry behavior on failure	”It handles it automatically” with no explanation of the path
Scope of action	Does it only read, or does it also send, pay, delete?	Permission scopes, list of integrations	Read and write permissions bundled together as one ask
Stopping points	When does it ask a human, and how does it roll back on failure?	Approval policy, audit logs4, rollback procedure	Approval step is optional and the default is allow-all

Bring this table with you and the questions in vendor meetings change. Instead of “is this an agent,” you end up asking “what’s the default setting.” Every vendor says yes to the first question. The answers split on the second.

Where the answers split across the three axes is also predictable. Ask about autonomy and vendors tend to answer generously — the more autonomous it sounds, the better it looks. Ask about scope of action and the answers get shorter; quite often they can’t even pull up the permission-scope documentation on the spot. Ask about stopping points and roadmap talk tends to follow. If the answer is “we’re adding audit logs next quarter,” that product belongs in a pilot, not a rollout.

Fewer vendors than you’d expect can answer all three axes with actual documentation. And that small group is the real shortlist for a comparison table. What narrows down the recommendation list isn’t a feature comparison — it’s this step of checking the documentation.

Watching Whether Permissions and Contract Terms Get Specific

Last time, I talked about how every AI that has actually landed with consumers was an AI with its name erased. This observation is one step earlier than that. Before a name gets erased, the market first has to agree on what goes underneath that name.

I don’t think a category matures the moment its definition sentence gets more precise. It matures when the permission-setting screens and the contracts get more precise.

That’s what happened with cloud. Early on, “what is cloud” was the search term, but almost no buyer asks that question now. Instead, they ask about regions, permission policies, availability guarantees. The question didn’t disappear — it moved. The moment abstract category questions shifted into concrete contract-term questions was that market’s point of maturity.

This reading can be turned into something falsifiable. If my judgment is right, “what does agent mean” shouldn’t just decline going forward — it should relocate to different search terms. The indicator is whether operational questions like “agent permission settings,” “approval workflow design,” “agent audit logs” fill that vacated space. Conversely, if definition searches quietly decline while operational questions don’t rise, that’s not maturity — that’s just interest cooling off. I’ll check the same data again in a few months.

The same logic holds for sellers, too. If you strike the word “agent” from your product page and replace it with exactly what tasks it can be trusted with and where it stops, buyers have less reason to go looking elsewhere for a definition. A visitor who leaves the page to go verify a definition elsewhere may abandon the product review altogether. Content that explains a category name contributes to the whole category — not to any one specific product.

Closing

Let me sum this up in three lines.

• In this model, meaning being placed ahead of recommend, I read as a signal that the model wants to check the comparison criteria first. That said, do keep in mind that search frequency itself is a bit higher for recommend.
• Right now, “agent” isn’t a product category — it’s a label with several layers mixed together. Even Anthropic offers its own definition while noting there’s no industry consensus.
• So we need to reverse the order of judgment. Before asking what it’ll recommend, ask what it can be trusted to handle.

Next time you get a briefing document reviewing agent adoption, try adding three more columns next to the product-name column: the scope it can choose on its own, the actions that change external state, and where approval and rollback sit. A product with all three columns blank isn’t a comparison candidate yet — it’s still a pre-candidate.

And in the review meeting, ask this just once: “If this tool executes something wrong, which one of us can undo it, and within how many minutes?” Once that question comes back with both a name and a time attached, that’s when the recommendation list starts to actually mean something.

References & Further Reading

Primary sources

• Anthropic, “Building effective agents”, 2024. 12. 19. Link ··· This is where the line between workflows and agents gets drawn. Read the definition paragraph up front before you get to the pattern descriptions.
• Anthropic, “Trustworthy agents in practice”, 2026. 4. 9. Link ··· It lays out the operational definition of the plan-act-observe-revise loop, alongside per-tool allow/approve/block settings. The three axes in today’s piece came from here.
• Google Developers Blog, “Developer’s Guide to AI Agent Protocols”, 2026. 3. Link ··· It maps out which layer each of MCP, A2A, UCP, AP2, A2UI, and AG-UI is responsible for. It’s the fastest way to confirm that this isn’t a single category.

Related past issues

• A technology that never comes up in the group chat isn’t a product ··· This one looked at why agents haven’t spread yet, through the lens of verification burden.
• AI agents are already ready — we’re the ones who haven’t caught up ··· This issue looked at autonomy, approval rates, and the share of irreversible actions using actual usage data.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Glossary

Footnotes

1. Workflow: Automation in which a human pre-defines the order in which tools and models get called. Because the path is fixed before execution, it’s easy to test and predict. ↩

2. MCP (Model Context Protocol): An open standard that standardizes how AI accesses external tools or data. It removes the need to write separate connector code for every single tool. ↩

3. AP2 (Agent Payments Protocol): A protocol that leaves verifiable evidence of who approved a payment made by an agent, and under what limit. It’s not about what gets bought, but about who gave permission. ↩

4. Audit log: A record of what action a system took, when, and under whose authority. It’s the evidence base for finding the rollback point when something goes wrong. ↩