StrongMocha published a vendor audit on June 9 titled “The Agent Trap: Why 90% of AI ‘Launches’ Are Infrastructure Liars,” arguing that the word “agent” has been stripped of technical meaning in enterprise software marketing. The piece, authored by Thorsten Meyer, proposes a five-question procurement filter and claims that only 10% of products currently marketed as agents qualify as actual autonomous infrastructure.

The Five-Point Filter

StrongMocha’s framework tests vendor claims against five questions. A product that fails three or more is “a feature, not infrastructure”:

  1. Does it run when no human is logged in? Real agents run on a schedule, trigger, or daemon. If it only works when a user opens a tab, it is a feature.
  2. Can you swap the model without losing the work? Real agents treat the model as substitutable. Runbooks, tools, memory, and workflows survive a model change.
  3. Where does the state live? Real agents persist state to a customer-controlled store with a queryable schema. Features persist to “your conversation history” inside the vendor’s database.
  4. What does the audit trail look like to your SOC? Real agents emit events into a SIEM or webhook stream. Features emit nothing, or vendor-side logs that cannot be ingested.
  5. What do you keep when the contract ends? Real agents leave exportable skills, prompts, runbooks, memory, and integrations. Features leave nothing but the labor sunk into the vendor’s UI.

The 90/10 Split

Meyer’s central claim: 90% of 2026 AI “agent” product launches are “chat boxes wired to SaaS via OAuth” with per-seat pricing, vendor-cloud-only deployment, conversation context as state, no SOC-ingestible audit trail, and nothing exportable at contract termination. The remaining 10% are “runtime, model-substitutable, governable” with per-action pricing, customer-controlled state, SIEM-emitting audit, portable skills, and the ability to survive a vendor change.

The piece names OpenClaw, Hermes, and CrewAI as examples of frameworks that pass all five filters: processes that operate continuously, maintain persistent state, and can be governed externally. It does not name specific vendors that fail the filters, instead describing a composite: “$30 per seat per month, targeting 4,000 paid users by year-end” that is “merely a chat feature integrated with existing SaaS tools.”

The Procurement Implication

The audit’s practical value is in procurement language. Meyer argues that the distinction between “feature” and “infrastructure” is now “a critical procurement skill” because pricing models diverge dramatically. Real agent infrastructure moves toward per-action pricing (60-85% cost savings over per-seat models, according to the piece), while feature-labeled products maintain traditional SaaS per-seat structures.

For enterprise buyers evaluating agent products, the five-question filter provides a concrete due-diligence checklist before signing purchase orders. The test is mechanical: run through the questions, count the failures, price accordingly.

Where the Framework Falls Short

StrongMocha’s filter effectively identifies infrastructure maturity but does not address operational security. A product can pass all five questions (runs autonomously, model-substitutable, customer-controlled state, SIEM-emitting, exportable) and still fail at basic trust boundaries. Varonis’s phishing research, published the same day, demonstrates this precisely: an OpenClaw agent passes every infrastructure filter but hands credentials to phishing emails because it lacks identity verification at the social layer.

The five-question filter is necessary but not sufficient. Enterprise procurement teams need a sixth question: “How does the agent verify the identity and authority of the humans and systems that request actions from it?” Infrastructure without trust governance is a well-audited system that can still be socially engineered.

That said, StrongMocha’s contribution is timing the right framework to the right market moment. The agent label inflation problem is real. Buyers need vocabulary to push back on vendors using “agent” as a premium-pricing label for chat features. The five-point filter gives procurement teams that vocabulary.