Agentic

Agency emerges from the model itself, not from pre-coded logic. True agentic behavior is driven by the LLM’s reasoning, not rigid predefined workflows.

Agency — the ability to perceive, reason, and act — comes from model training, not from external code orchestration. But a working agent product needs both the model and the harness. The model is the driver, the harness is the vehicle.

At the core of every agent is a neural network — a Transformer, an RNN, a learned function — that has been trained, through billions of gradient updates on action-sequence data, to perceive an environment, reason about goals, and take actions. Agency is never granted by the surrounding code. It is learned by the model during training.

For example, given enough context and access to the right tools, the model can drive the task forward by deciding what to inspect, what action to take next, and how to revise its plan based on feedback. The harness provides the interface and constraints, but the step-by-step behavior is generated by the model rather than pre-written as a fixed workflow.

This does not mean that every agent product should pursue full autonomy. In most real-world applications, the model may perform local perception, reasoning, and action selection, but its operating space is still bounded by the harness: what it can see, which tools it can call, what actions it is allowed to execute, and when human approval is required. In other words, the model can exhibit agentic behavior without being given unlimited autonomy. The real design question is not whether to “add agency” through code, but how much freedom the trained model should have inside a controlled environment. In low-risk creative or productivity tasks, the system can give the model more room to explore and adapt. In high-stakes domains such as financial trading, legal work, or medical decision-making, the same agency must be narrowed through validation, permissions, guardrails, and human oversight.