I stopped my AI agents from being expensive black boxes (here is the architecture)

Description and use

The workflow‑centric architecture cuts costs by using small, scoped agents, provides per‑step metrics for auditability, and allows editing of the graph to swap models or budgets.

Problem

Single "do-everything" AI agents are expensive, unpredictable, and opaque, making it hard to control costs and audit actions.

Outcome

The workflow‑centric architecture cuts costs by using small, scoped agents, provides per‑step metrics for auditability, and allows editing of the graph to swap models or budgets.

Workflow

• Define job in chat
• LiveContext builds a workflow graph
• Schedule trigger (e.g., daily)
• Fetch emails
• Agent step: classify emails (small model, limited budget)
• Decision node: simple vs. angry
• Agent step: draft reply (small budget)
• Human‑approval step for angry emails
• Approval screen app for teammates
• Execute workflow with scoped agents

Evidence

A single "do‑everything" agent is expensive, unpredictable, and a black box.

Source

In LiveContext the workflow is the top of the stack, and agents are steps inside it.

Source

Every agent step has a credit budget it cannot exceed. When the budget is gone, the step stops.

Source