## Sources

1. [The Subsidy Ended: What Tool-Using Agents Actually Cost](https://www.oreilly.com/radar/the-subsidy-ended-what-tool-using-agents-actually-cost/)

---

### **The Subsidy Ended: What Tool-Using Agents Actually Cost** — By Bennie Haelen

**Main Arguments**
*   **Visibility of Existing Costs:** The transition to usage-based billing for AI agents—specifically highlighted by GitHub Copilot’s June 1, 2026, policy change—did not actually increase the cost of agentic work; it simply made previously subsidized costs visible to the user [1, 2].
*   **The "Loop" is the Expense:** Unlike simple chat completions, tool-using agents operate in loops (reasoning, calling tools, and processing results), with each iteration carrying forward an expanding context window that drives up token consumption and costs [2, 3].
*   **Architecture Over Budgeting:** Managing AI costs should be treated as an **architectural governance problem** rather than a simple budgeting issue. Budget caps merely report overspending after it occurs, whereas architectural guardrails prevent it by design [4, 5].
*   **Industry-Wide Shift:** While GitHub is the current focus due to its recent billing changes, the industry is moving toward usage-based models for all agentic workloads as the era of flat-rate subsidies for autonomous AI ends [6, 7].

**Key Takeaways**
*   **Usage-Based Billing Mechanics:** Under new plans like GitHub Copilot Pro, users receive a monthly pool of AI credits (priced at one cent each) that are consumed based on the specific model used and the volume of tokens processed [1].
*   **The Cost of Vagueness:** Agent costs do not scale linearly; they scale with the number of turns required to complete a task. Vague prompts force agents into "discovery loops," which are significantly more expensive than well-defined, bounded queries [3, 8].
*   **Tool Design as Cost Discipline:** The size and complexity of a tool catalog now directly impact the bill. Verbose tool descriptions and bloated catalogs consume "rent" in the context window during every turn of an agent’s loop [9].
*   **Platform-Level Controls:** Effective cost governance belongs in a **control plane** above the agent. This layer should enforce rules such as routing tasks to cheaper models, bounding the number of iterations, and capping tool-result payloads [5, 10].

**Important Details**
*   **Context Carryover:** In many agent harnesses, every pass through a loop carries forward prior messages, tool descriptions, and retrieved files. Even when cached, this state preservation constitutes metered work [3].
*   **Prompting Patterns for Savings:**
    *   **Specific Inputs:** Users should prompt agents like a "short requirement" (naming specific tables, fields, and date ranges) rather than asking broad questions. This "collapses the loop" by providing necessary information on the first turn [8, 11].
    *   **Plain Text Outputs:** Requesting Markdown or plain text for intermediate steps saves significant costs, as rich HTML formatting is expensive to generate and often unnecessary until the final deliverable is confirmed [11].
*   **Guardrails vs. Prompts:** While individual prompt discipline can reduce costs, it is fragile because it relies on consistent human behavior. **Guardrails** at the platform level make cheaper and safer behavior the default, regardless of the user's specific prompt [10].
*   **Strategic Response:** Organizations that view these changes as mere "pricing events" will only perform minor optimizations, while those that view them as "architecture signals" will build durable, platform-based cost controls that remain effective regardless of vendor changes [7].