Challenges and Agent Behavior in Collaborative Swarm Environments.

Bigger models tend to perform better for a specific task in a collaborative swarm environment. This could be because the model's training is optimized for efficiency and task completion. This could mean that agents will choose shorter execution path most of the time in favor of performance.

Guidance is not equal to Guarantees

One of the roles that emerged in the Agentic era was Prompt Engineer, a professional that understood the complexity of agent reasoning, and could provide clear instruction to the agents to follow, to to reduce the impact of non-deterministic behavior of the agents, providing precise, logical answers and aligned to the objective.

Today we use prompts in every aspect of the interaction with agents, from a simple request to explain tools utilization rules, inter-agents interaction, code reviews and much more. We fill repositories with Markdown files expecting the agent to comply as requested. We can quickly become trapped in a cycle of tuning all these prompts, because the final results, or the way tools are used, often end up far from what was expected.

As humans, we naturally try to transfer the same rule-based systems we use to govern human behavior to agents, but guidance is not equal to guarantees, LLMs are not guaranteed to follow our set of prompts --rules-- because: LLMs are trained to predict and complete the most probable path toward an inferred objective, not to enforce operator intent. With this in mind, when building agents, it is important to consider that: as capability increases, agents tend to optimize for inferred task completion rather than strict instruction compliance, will prevent undesirable or unpredictable behavior.

This raises a broader question. If governance mechanisms are repeatedly compensating for undesirable behavior, are we addressing the root cause? Or are we attempting to govern behavior that originated from a capability mismatch?

Governance is only one layer

The typical response to unexpected behavior is governance.

These mechanisms can certainly improve outcomes. However, they often operate after capability selection has already occurred.

In many cases we attempt to govern behavior without first understanding whether the selected capabilities are appropriate for the task, and this creates an important architectural question. Are all governance problems actually governance problems? Or are some of them capability-assignment problems?

One observation emerging from my experimentation is that we tend to place models at the center of system design. The question usually becomes: Which model should we use?

A more useful question may be: What capabilities does the task require?

Each characteristic can be assigned a weight, these weights can be normalized to generate a capability profile for the task. Separately, models can be characterized using the same dimensions, the result is a capability profile for each available model.

Instead of selecting models based on popularity, benchmark scores, or intuition, we can compare capability profiles against task requirements.

Task Profile ↔ Capability Profile

The purpose of these mechanisms is no longer arbitrary governance, their purpose is to reinforce specific missing capabilities.

Prompts express intent; compliance depends on inference, and inference is only partially observable. As agentic systems become increasingly autonomous, prompt-based governance becomes less capable of providing behavioral guarantees on its own.

This doesn’t mean we should drop prompts. It means prompts should be understood for what they are: A guidance mechanism rather than an enforcement mechanism, and governance is still necessary.

But before building governance layers, we need to correctly characterize the task and assign the appropriate capabilities.