Glages: The Agent Is a Replaceable Mechanism
AI products are often designed around the current capabilities of a particular model. The prompts, tools, memory, and orchestration become the center of the architecture.
Glages places the formal business model at the center instead.
The agent becomes one execution mechanism among several.
The Activity Must Outlast the Agent
A payment, refund, approval, or delivery process should retain the same meaning when the execution technology changes.
A company may replace one LLM with another. A bounded classification task may later become deterministic code. A human-approved step may eventually be automated. An external provider may change.
If the business process is stored mainly in prompts and agent behavior, each change risks altering the activity itself.
When the process is defined by a formal model, the mechanism can change inside a stable boundary.
Approved Typed Input
The agent does not receive the entire organizational environment and an instruction to determine what matters.
It operates only inside the world expressed by its approved typed Input.
The Input can contain the relevant facts, references, evidence, and explicitly available alternatives for one task. This reduces the need for the agent to reconstruct the process and limits the consequences of an incorrect interpretation.
The agent may interpret bounded text, extract structured facts, compare evidence, draft content, or choose among modeled options. It may not silently add an action or assume a missing fact.
The Model Defines Authority
The model determines required facts, permitted actions, allowed changes, valid results, and stopping conditions.
The agent may help resolve whether a customer message refers to one of three known orders. It does not decide that an order may be refunded when the modeled conditions are not satisfied.
The agent may draft an explanation of a declined payment. It does not convert the decline into approval because another action appears more helpful.
Controlled Substitution
Because the task boundary is formal, execution mechanisms can be substituted and compared.
A deterministic parser may handle a structured input. A small language model may classify a bounded phrase. A larger model may be used only when the evidence is complex. A human may remain responsible where authority or unresolved judgment is required.
The system can choose the mechanism appropriate to the task without redefining the process.
Architectural Durability
This architecture is less dependent on the rise and fall of individual AI products.
LLMs will improve, become cheaper, and change their interfaces. Some tasks that require language models today may become ordinary software tomorrow. New execution mechanisms will appear.
The model remains useful because it preserves the meaning of the activity rather than the behavior of one tool.
The agent is valuable, but replaceable. The formal model is the durable asset.