10 Things You Need to Know About Token Efficiency in GitHub Agentic Workflows
Agentic workflows are transforming repository maintenance, but their token consumption can spiral out of control. In this listicle, we share ten essential insights from our journey to optimize token usage in GitHub’s own automated workflows. From logging methods to daily auditing and optimization agents, these tips will help you keep costs in check while maintaining code quality.
1. What Are Agentic Workflows?
Think of agentic workflows as a fleet of digital street sweepers: each one tidies up small messes in your repository automatically. These agents handle tasks like dependency updates, code linting, and minor refactors, running as part of your CI pipeline. They dramatically improve repo hygiene and developer productivity by reducing manual, repetitive work. However, because they are triggered automatically—often on every push or schedule—they can rack up API costs without anyone noticing. Understanding what agentic workflows do is the first step to controlling their token budget.
2. The Stealthy Cost of Automation
Unlike interactive developer sessions where you see each API call, agentic workflows operate in the background. They are triggered by events (commits, PRs, etc.) and run without human oversight. This “out of sight, out of mind” nature means token usage can accumulate rapidly, especially when workflows are poorly configured. Because each execution is fully automated, costs can escalate before you realize there’s a problem. Recognizing this hidden cost is crucial for teams adopting agentic automation at scale.
3. Why Workflows Are Easier to Optimize Than Interactive Sessions
Session-based work is unpredictable—developers explore, debug, and experiment. In contrast, agentic workflows have their entire logic defined in a YAML file. Every step, condition, and API call is specified upfront. This deterministic nature makes it much simpler to analyze and optimize. You can inspect the workflow, understand exactly what it does, and identify inefficiencies without guesswork. That’s why we focused our optimization efforts on agentic workflows rather than human coding sessions.
4. The Foundation: Logging Token Usage
Before you can optimize, you must measure. The first hurdle was that different agent frameworks—Claude CLI, Copilot CLI, Codex CLI—each emit logs in their own format. Historical runs often had incomplete data. To get a unified view, we needed a consistent way to capture token consumption across all runs. This led us to a centralized approach that normalizes data regardless of the underlying agent.
5. The API Proxy: A Centralized Monitoring Point
Our security architecture already required all agentic workflows to communicate through an API proxy. This proxy sits between the agent and the model providers, preventing direct credential access. It turned out to be the perfect vantage point for monitoring: we could intercept every API call, log input and output tokens, cache hits, model used, and timestamps. By leveraging this existing component, we avoided adding new instrumentation to each framework. The proxy now produces a token-usage.jsonl artifact per workflow run.
6. Granular Token Data at Your Fingertips
Each record in the JSONL file contains: input tokens, output tokens, cache-read tokens, cache-write tokens, model identifier, provider, and timestamps. Combined with workflow logs, this data gives a complete picture of how tokens are spent over time. For example, you can see which steps consume the most tokens, whether caching is effective, and how model choices affect costs. This granular visibility is essential for informed optimization decisions.
7. Daily Token Usage Auditor
With token data streaming in, we built an automated auditor that runs daily. It reads recent token-usage.jsonl artifacts, aggregates consumption per workflow, and generates a structured report. The auditor’s job is to flag workflows that have significantly increased token usage compared to their baseline, highlight the most expensive workflows, and detect anomalous runs (e.g., a workflow that normally takes 4 LLM turns suddenly uses 18). This proactive monitoring lets us catch regressions early.
8. Daily Token Optimizer: Finding the Fixes
When the auditor flags a workflow, a second agent—the Daily Token Optimizer—kicks in. It analyzes the workflow’s source code and recent logs to identify concrete inefficiencies. Then it creates a GitHub issue describing the problem and suggesting specific optimizations. For example, it might propose reducing the number of LLM turns, caching intermediate results, or switching to a cheaper model. The Optimizer has uncovered many subtle issues we would have missed manually.
9. Common Inefficiencies Discovered
Through this process, we found several recurring patterns: workflows that made unnecessary repeated calls with identical contexts, excessive tool usage, and suboptimal prompt structures. One standout case was a workflow that ran 18 LLM turns when 4 would suffice simply because it didn’t pass context efficiently. Another had a loop that re-initialized the agent state each iteration, wasting tokens. These insights led to immediate fixes and guidelines for future workflow authors.
10. Preliminary Results and Next Steps
After a month of active optimization, we saw a 30% reduction in total token consumption across our critical workflows—without sacrificing quality. The auditor and optimizer themselves are agentic workflows, so they also contributed to token usage, but the savings far outweighed the cost. Moving forward, we plan to extend the system to automatically apply optimizations via PRs and to share best practices with the community. Token efficiency is an ongoing journey, but the tools we’ve built make it manageable.
Agentic workflows are powerful allies in repository maintenance, but they demand careful stewardship. By measuring, auditing, and optimizing, you can keep costs under control while reaping the benefits of automation. Start with logging, then build feedback loops—your token budget will thank you.