note №.005 · 2026 · 03 · 157 min - or one elevator argument

Why agentic is quietly
becoming useless.

A small vocabulary for replacing agentic theatre with an inspectable description of system behavior.

A useful word is losing its meaning, not because of anyone in particular, but because everyone is using it for everything.

An autocomplete is agentic. A chatbot is agentic. A fixed workflow with one model call is agentic. A system that creates accounts and deploys code is also agentic.

When one word covers all four, it stops helping engineers, buyers, and users understand risk.

The problem is not the word.

"Agentic" names a real shift: software can increasingly interpret goals, choose intermediate steps, use tools, observe results, and continue without every transition being specified in advance.

The problem is using the label where a behavioral description is needed.

In a product review, architecture document, or sales conversation, I would replace "agentic" with answers to seven questions.

1. What goal can the system pursue?

Is the goal a narrow task such as enriching one alert, or an open-ended outcome such as investigating suspicious identity activity?

Who supplies the goal? Can the system rewrite it? What prevents retrieved text or a tool response from changing it?

Goal scope determines both capability and attack surface.

2. What can it decide?

Some systems select among fixed branches. Others generate a plan, add steps, delegate to another agent, or decide that more evidence is needed.

Describe the decision:

  • choose a source;
  • rank hypotheses;
  • select a tool;
  • request approval;
  • stop or continue;
  • propose an action.

This is more informative than an autonomy label.

3. Which tools can it use?

Tools turn language into effects.

List whether they read data, write state, execute code, spend money, or affect a person. Explain how arguments are validated, credentials are scoped, and risky actions are approved.

OWASP's guidance on agentic threats focuses heavily on tool misuse, identity, memory, and cascading behavior because authority changes the consequence of a model error.

4. What state does it retain?

Does the system remember only the current request, maintain workflow state, or write durable memory that affects future decisions?

State needs provenance, ownership, correction, expiry, and deletion. "The agent learns" is not a sufficient data model.

5. What feedback can it observe?

Can it inspect a tool result, test whether an action succeeded, compare evidence, or see a human correction?

Without observation, a sequence of model calls may look autonomous while being unable to detect its own failure.

6. What authority does it borrow?

An agent rarely has authority of its own. It borrows authority from a user, service account, integration, or organization.

The architecture should bind that authority to:

  • principal;
  • tenant;
  • task;
  • tool;
  • target;
  • duration;
  • approval.

This is where product language becomes security architecture.

7. How does it stop?

Every system needs stopping conditions:

  • goal completed;
  • evidence insufficient;
  • budget reached;
  • policy denied;
  • approval expired;
  • repeated failure;
  • human intervention;
  • risk threshold exceeded.

An agent that can continue but cannot explain when to stop is an unbounded loop with branding.

A more useful description.

Instead of:

We built an agentic SOC analyst.

Try:

The system investigates one alert at a time. It can retrieve evidence from three approved sources, form and test hypotheses, and draft a disposition. It cannot execute response actions. It stops when evidence is insufficient, its step budget is reached, or an analyst intervenes.

That description is less exciting and much easier to evaluate.

Autonomy is multidimensional.

A system can have high planning freedom but read-only tools. It can have fixed planning but highly privileged actions. It can have durable memory but no tools.

For each workflow, map:

DimensionLowHigh
goal scopefixed taskopen-ended outcome
planningfixed pathgenerated and revised plan
toolsread-onlyprivileged effects
memoryrequest onlydurable shared memory
supervisionevery stepexception only
reversibilityeasy rollbackirreversible consequence

This makes architecture and governance discussions concrete.

Keep the word, add the verbs.

I still use "agentic" as shorthand. It is useful for naming a family of systems.

But the word should begin the conversation, not finish it.

Tell me what the system observes, decides, retrieves, remembers, changes, and refuses to do. Then tell me who gave it permission.

That is the product.

Sources and further reading.

filed under →aiopinionsagents
↬ read next:

Rethinking SecOps: designing an AI-native security operations platform.

Traditional security platforms were built around dashboards and queues. AI-native SecOps platforms need evidence, workflow memory, typed actions, guardrails, and analyst control.

continue →