Your Agentic AI Is Recreating the Meetings It Was Supposed to Replace

Jeremy McEntire, Head of Engineering at Wander, explains how agentic AI is recreating the meetings it was supposed to replace, and what companies can do about it. This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.

“No one of us is as dumb as all of us.” — Demotivational poster, author unknown

We ran a controlled experiment. Four AI architectures—single agent, hierarchical multi-agent, stigmergic multi-agent, and pipeline—were given identical software engineering tasks. Same compute budget. Same instructions. Same goal. The single agent completed 28 out of 28 tasks, while the pipeline completed zero. It consumed its entire $50 compute budget on planning and produced no deployable code. No humans were involved. The agents held meetings, reviewed each other’s work, rejected 87 percent of submissions—and shipped nothing.

The pipeline didn’t fail because of bad prompts or the wrong model. It failed because it was subject to the same structural forces that make human bureaucracies dysfunctional. We’ve spent decades building organizations that talk about work rather than do it. When we trained AI on the outputs of those organizations, it learned to do the same. This is not a bug you can patch.

Why This Happens

Every coordinating system—human or artificial—must compress information to function at scale. A CEO can’t make every decision in the company, and an AI pipeline can’t transmit every detail between agents without loss. So both compress: summaries, status updates, key judgments, and formatted outputs. Compression creates gaps. What the summary captures becomes visible. What it discards becomes invisible—not just unread, but inconceivable within the compressed frame.

Selection operates on the compressed representation. In human organizations, the signals that survive are the ones that are easy to transmit, politically safe, and consistent with what leadership already believes. Truth is orthogonal to fitness. Sometimes truth survives. Often it doesn’t.

In AI pipelines, the same dynamic plays out. Each agent receives a compressed version of what came before, optimizes for the approval of the next agent in the chain, and passes along a further-compressed version of its work. The review stages reject 87 percent of submissions, not because the work is bad, but because the channel has degraded to the point where nothing fits.

I call this dysmemic pressure—when a system consistently selects for messages that are easy to transmit over messages that are true. Think of a farmer who eats the best tomatoes and plants the rest. Eventually, the crop is terrible. Geneticists call that dysgenic pressure; dysmemic borrows the same logic, applied to ideas instead of genes. Organizations under dysmemic pressure don’t fail because their people become dumb. They fail because they systematically cultivate the wrong crop. AI systems trained on those organizations learn the cultivation pattern, not the tomatoes.

This is the same mechanism I spent several years formalizing in The Cage and the Mirror—the structural forces that trap intelligent organizations in self-reinforcing patterns of dysfunction. The AI finding isn’t a separate problem. It’s the same cage, built faster, at lower cost, and deployed at scale before anyone asked whether we wanted to replicate it.

The Part Nobody Told You

Here is the finding that changes the calculus: the same mechanism that produces organizational dysfunction produces jazz. This is not a metaphor. It is a formal result.

When Miles Davis assembled his Second Great Quintet—musicians with deliberately incompatible approaches, minimal direction, no preset harmonic schemes—he was engineering a lossy channel. The gap between what each musician intended and what the others perceived was the generative residual. Herbie Hancock played what he thought was the wrong chord. Davis incorporated it into a new direction. Neither musician would have composed what emerged from the collision of their partial views.

The Nokia middle manager and the jazz musician are running the same process. Information enters a lossy channel. The channel strips something. The receiver reconstructs using their own context, priors, and judgment. When the selection environment rewards fit with the sender’s preferences—organizational fitness over external accuracy—the result is dysfunction. When it rewards functional novelty, the result is creation.

Same physics. Different pressure.

This means you cannot eliminate the mechanism that produces AI dysfunction without eliminating the mechanism that produces AI creativity. They share a root cause. The design problem is not elimination. It is tuning.

We proved this tonight. Large AI models appear to lose specialized knowledge as they scale—but they haven’t actually lost it. It’s still in there, buried under overlapping information that the standard tools can no longer distinguish. Adding a small, calibrated amount of randomness—like adding just enough static to hear a radio station more clearly—recovers it almost completely. A system that was performing correctly on specialized tasks 8.7 percent of the time jumped to 99.3 percent with the right amount of noise. Too little noise and the signal stays buried. Too much and you get chaos. The sweet spot is real, measurable, and predictable. The theory told us exactly where to look. It was right.

AI systems don’t lose specialized knowledge as they scale. They just lose the ability to find it. We found the key.

Five Conditions That Predict Which Way It Goes

Our research identifies five checkable properties that determine whether an AI system—or a human organization—will drift toward dysfunction or toward genuine productivity. When all five are met, calibrated variance produces a net benefit. When any fails, dysfunction is the default.

The conditions are:

• The system must be operating below its potential (there must be room to improve);
• The integration function must be nonlinear (rigid pass-throughs cannot benefit from variance)
• The variance must be able to reach the improvement region (noise that can’t touch the problem can’t fix it)
• The gain at the higher level must outweigh the cost at the lower level
• The lower level must degrade gracefully rather than catastrophically under pressure.

These conditions apply equally to your AI deployment and your human organization. Most corporate transformation initiatives fail—the research puts that failure rate at 60-80 percent—because the reform proposal enters the same selection environment it’s trying to change. The channel filters it. What survives is vocabulary-only reform that changes the language without changing the pressure. If your organization is under dysmemic pressure, adding an AI layer won’t relieve it. It amplifies it.

Check these five conditions against your current agentic AI deployment. If your pipeline architecture has brittle handoffs between agents—where one agent’s failure cascades immediately into the next—you have violated the fifth condition. Add redundancy. Build graceful degradation into the handoffs.

If your agents are optimizing for each other’s approval rather than for the end goal, you have the wrong selection pressure. Redirect the reward signal toward outcomes rather than process compliance. If your coordination layer is more complex than your execution layer, you have an inverted architecture. Our single agent—one clear task specification, high internal flexibility—outperformed every multi-agent configuration. Coordination complexity is a cost, not a feature.

What To Do Next

The question for every organization deploying agentic AI is not whether the technology will replicate your dysfunction. It will. The question is whether you are designing the selection environment to channel that toward creativity or toward bureaucratic collapse. Three immediate steps:

• First, measure output-to-planning ratios in your current deployments. If your agents are spending more compute on coordination than execution, you have a pipeline problem, not a technology problem.
• Second, simplify before you scale. The single-agent result is not an argument against multi-agent systems—it is an argument against premature coordination complexity. Start with the simplest architecture that accomplishes the task. Add coordination only when you have evidence that it helps.
• Third, check your selection pressure. What does your system reward? If it rewards process adherence, you will get process. If it rewards outcomes, you have a chance at results.

The same diagnosis applies to the humans in the room. Dysmemic pressure doesn’t care whether the channel is made of people or parameters. AI systems don’t fail because they are unintelligent. They fail for the same reasons intelligent organizations fail: because the channel between intent and output has degraded past the point where a useful signal survives.

You cannot engineer your way out of that by switching vendors or tweaking prompts. But you can design for it. The physics are the same. The choice of selection environment is yours.