IV.
[deferred to copy pass]
Agents promise what the market wants: autonomous action at scale. Systems that do not merely respond but initiate. That do not require constant instruction but pursue goals across multiple steps. That act.
The appeal is real. Over 60% of organizations are experimenting with agents. The pull toward autonomous execution is strong because execution is what the market has learned to value.
But agents inherit the architecture's core flaw—execution without interpretation—and they multiply it. They scale action. They do not scale interpretation. The same absence that produces drift in simpler systems produces compounding failure when systems act autonomously across many steps.
The experimentation is broad. Nearly two-thirds of organizations are testing agents in some capacity. Almost a quarter claim to be scaling agentic AI somewhere in their enterprise.
The numbers become less impressive under scrutiny. Most organizations scaling agents are doing so in only one or two functions. In any given business function, no more than 10% report scaling. The pattern is familiar: widespread experimentation, narrow actual deployment.
The infrastructure for agentic deployment does not exist in most organizations.
Over 40% are still developing their agentic strategy. More than a third have no formal strategy at all. Investment is happening without architecture. Organizations are deploying agents into environments that were never designed for autonomous action, without the interpretive infrastructure that would make autonomy governable.
The enthusiasm outpaces the readiness. This is not new. It is the same pattern from earlier adoption waves, accelerated.
The market is saturated with systems labeled as agents that are not agents in any meaningful sense.
Agent washing is the rebranding of existing products—robotic process automation, chatbots, AI assistants—as agents, without substantial agentic capabilities. Existing scripts and workflows are relabeled without true autonomy, without decision boundaries, without accountability structures.
The test is simple: if a system cannot safely act, adapt, and recover without constant human oversight, it is not agentic. It is automation with better marketing.
Of the thousands of vendors claiming agentic AI capabilities, perhaps 130 are building something real.
The confusion is not merely noise. It obscures the structural question: what does autonomy actually require?
Most use cases positioned as agentic today do not require agentic implementations. Current models lack the maturity to autonomously achieve complex business goals or follow nuanced instructions over extended interactions. The capability demonstrated in controlled environments does not survive contact with production complexity.
The market is buying a category that does not yet exist in the form being sold.
The reliability mathematics of multi-step execution are unforgiving.
Assume each step in an agent workflow succeeds 95% of the time. That sounds robust. A twenty-step workflow, at that rate, has approximately a 36% chance of completing without error.
Production systems for critical business processes typically require 99.9% reliability or higher. The gap between agent capability and production requirements is not incremental. It is structural.
Agents can initiate action. They rarely complete workflows to the standards that regulated businesses require. The gap between action and completion is what sinks these programs.
Agent failures have specific signatures.
Instruction-following deviation: the agent diverges from intended behavior partway through execution. Long-range contextual misuse: information from earlier steps is misapplied in later ones. Sub-intention errors: omission of necessary steps, redundancy of completed steps, disordering of sequences.
Each step proceeds without validating that the previous step was correctly interpreted. The agent cannot question its own premises. Multi-step execution without interpretive checkpoints means errors compound rather than surface.
Organizations are discovering the gap in real time.
Agentic AI can act. It rarely finishes workflows in a way that meets business standards. Pilots that looked promising fail to translate into production-ready solutions. The obstacles are consistent: legacy systems that were never designed for agentic interactions, data architectures that create bottlenecks, technical debt that teams underestimated.
Most agents still rely on conventional APIs and data pipelines. These dependencies limit autonomous capability and create failure points that undermine the promise of self-directed execution.
The prediction: over 40% of agentic AI projects will be canceled by 2027. The causes cited are familiar—escalating costs, unclear business value, inadequate risk controls. The same causes that drive non-agentic AI abandonment, but accelerated by the compounding dynamics of multi-step autonomous execution.
Agents are not a solution to the interpretation problem. They are an acceleration of execution without interpretation.
The mirage is capability that looks like understanding. Systems that act with confidence across complex sequences, producing outputs that resemble goal-directed behavior. The appearance is compelling. The structure is absent.
The reality is action that scales faster than the ability to govern it. More agents, deployed more broadly, operating with the same structural absence that produced drift in simpler systems. The gap between what agents do and what situations require does not close with more capability. It widens.
The question is not whether agents can act. They can. The question is whether action without interpretation produces anything other than faster drift.
That question remains open.