Introduction: The Cost of Automated Repo Maintenance
GitHub Agentic Workflows act as automated caretakers, tidying up minor issues across your repository much like a team of street sweepers. These workflows significantly improve repository hygiene and code quality. However, as with any automated agent-driven system, token consumption costs are becoming a growing concern for developers. Because these CI jobs are triggered automatically and run on a schedule, expenses can accumulate silently, often going unnoticed until they impact budgets.
Fortunately, optimizing automation is often simpler than optimizing interactive developer sessions. Work performed during a developer's desktop session can be unpredictable, but agentic workflows have their entire execution defined in YAML, making each run predictable and repeatable.
Because we both maintain and use GitHub Agentic Workflows in our own repositories, token efficiency is as important to us as it is to our users. Starting in April 2026, we began a systematic effort to optimize token usage across many of our daily workflows. This article outlines our instrumentation process, the optimizations applied, and the preliminary results we've achieved.
Instrumenting Token Usage
Our repositories rely on hundreds of agentic workflows for continuous integration and maintenance. All workflows run as GitHub Actions and are subject to real API rate limits. As we develop these workflows, we are essentially building the plane while flying it, consuming tokens like jet fuel along the way.
Before we could optimize token consumption, we needed clear visibility into where tokens were being spent. The first challenge was that each agent framework—Claude CLI, Copilot CLI, Codex CLI—emitted logs in different formats, and usage data for historical runs was often incomplete. Fortunately, the security architecture of agentic workflows uses an API proxy to prevent agents from directly accessing authentication credentials. This proxy provided a unified point to capture token usage across all runs in a standardized format, regardless of the underlying agent framework.
Today, every workflow produces a token-usage.jsonl artifact, containing one record per API call. Each record includes input tokens, output tokens, cache-read tokens, cache-write tokens, model name, provider, and timestamps. By combining this data with the rest of the workflow logs, we gained a historical view of typical token expenditure, enabling us to optimize for future runs.
Workflows Optimizing Workflows
Armed with detailed token data, we built two daily optimization workflows to continually improve efficiency.
Daily Token Usage Auditor
The Daily Token Usage Auditor reads token usage artifacts from recent workflow runs, aggregates consumption by workflow, and posts a structured report. Its primary job is to flag any workflow that has significantly increased its recent token usage, highlight the most expensive workflows, and identify anomalous runs—for example, a workflow that normally completes in four LLM turns but suddenly takes 18 turns.
Daily Token Optimizer
When the Auditor flags a workflow, the Daily Token Optimizer steps in. It examines the flagged workflow's source code and recent logs, then creates a GitHub issue that describes concrete inefficiencies and proposes specific optimizations. The Optimizer has uncovered many inefficiencies that we otherwise would have missed.
Interestingly, both the Auditor and Optimizer are themselves agentic workflows. They run daily, consuming tokens of their own—but their cost is far outweighed by the savings they generate across our entire workflow fleet.
Results and Future Directions
Preliminary results show a noticeable reduction in total token consumption across our most-used workflows. By identifying redundant API calls, optimizing prompt structures, and improving caching strategies, we've lowered costs without sacrificing functionality. We continue to refine these optimizations and plan to share more detailed benchmarks in a future post.
For developers looking to improve token efficiency in their own agentic workflows, we recommend starting with instrumentation—without data, optimization is guesswork. The API proxy approach used here is a practical pattern that can be adapted to other environments.