10 Ways to Slash Token Waste in GitHub Agentic Workflows

Agentic workflows are like having a tireless team of street sweepers keeping your repository clean. They automate small maintenance tasks, improve code quality, and catch issues before they grow. But there's a catch: every time these autonomous agents run, they consume tokens—and because they're automatically triggered, costs can quietly balloon without anyone noticing. GitHub faced this challenge head-on in April 2026, when we began systematically optimizing token usage across hundreds of workflows we rely on daily. This article distills our journey into ten actionable insights, from measuring token consumption to building self-optimizing systems. Whether you're using Claude CLI, Copilot CLI, or Codex CLI, these strategies will help you keep your agentic workflows lean, efficient, and cost-effective.

1. Understand Why Agentic Workflows Are a Prime Optimization Target

Unlike interactive developer sessions—where every keystroke is unpredictable—agentic workflows run deterministic, predefined steps captured in YAML. That repeatability makes them ideal for optimization. In a typical CI pipeline, a workflow might execute the same sequence of LLM calls dozens or hundreds of times a day. Each call consumes input tokens, output tokens, cache reads, and cache writes. Because the work is fully specified in code, you can analyze exactly where tokens are going and pinpoint waste. This is far harder with ad-hoc sessions, where human behavior introduces variability. By focusing on agentic workflows, you get the highest return on your optimization efforts.

10 Ways to Slash Token Waste in GitHub Agentic Workflows — Source: github.blog

2. Start with Comprehensive Token Logging

Before you can reduce token consumption, you need a clear picture of how tokens are used today. GitHub initially faced a challenge: different agent frameworks—Claude CLI, Copilot CLI, Codex CLI—all log token data in different formats. Worse, historical runs often had incomplete information. The solution was to leverage the API proxy that sits between agents and authentication credentials. This proxy, already part of the security architecture, captures every API call in a single, normalized format. Now every workflow outputs a token-usage.jsonl artifact containing input tokens, output tokens, cache-read tokens, cache-write tokens, model name, provider, and timestamps. This centralized log becomes the foundation for all optimization.

3. Build a Daily Token Usage Auditor

Once token logging is in place, set up a workflow to audit consumption daily. GitHub created a Daily Token Usage Auditor that reads token artifacts from recent runs, aggregates consumption by workflow, and posts a structured report. The auditor's job is to flag any workflow that shows a significant increase in token usage, highlight the most expensive workflows, and detect anomalous runs—for example, a workflow that normally completes in four LLM turns suddenly taking eighteen. Without an automated auditor, these spikes can go unnoticed until the monthly bill arrives. The auditor runs as an agentic workflow itself, so it's self-referential and self-improving.

4. Deploy a Daily Token Optimizer That Suggests Fixes

When the auditor flags a problem, a companion Daily Token Optimizer springs into action. It examines the flagged workflow's source code and recent logs, then creates a GitHub issue with concrete inefficiencies and specific optimization proposals. For example, it might notice redundant system prompts, overly long context windows, or unnecessary retries. The optimizer has caught many subtle inefficiencies that human developers would have missed because they don't scroll through hundreds of log lines. Like the auditor, the optimizer is itself an agentic workflow, creating a virtuous cycle of continuous improvement.

5. Use Cache Reads and Writes Wisely

Token logging revealed that cache operations—both reads and writes—are often underutilized. Many workflows re-request the same context repeatedly, ignoring cache-read tokens that could save input tokens. On the flip side, some workflows perform cache writes that are never used, wasting output tokens. By analyzing the token-usage.jsonl artifact, you can identify opportunities to increase cache hit rates. For instance, you can restructure prompts to reuse cached system messages or add cache_control directives. Even small improvements in cache efficiency compound across thousands of daily runs.

6. Right-Size Your Model Choices

Not every agentic task needs the most powerful (and most expensive) model. The token logs include the model field, enabling you to see which models are used in each workflow. GitHub found that many maintenance tasks—like formatting, linting, or low-level code reviews—could be handled by faster, cheaper models without sacrificing quality. By systematically downgrading model choices for less critical steps, some workflows cut token costs by 30-40%. The key is to experiment carefully: run a side-by-side comparison before switching and monitor downstream quality metrics.

7. Prune Unnecessary LLM Turns

Agentic workflows often involve multiple LLM calls per run. The token logs reveal how many turns each workflow takes, and the auditor flags outliers. In one case, a workflow that should have finished in four turns was taking 18 because of a circular reasoning loop. The optimizer identified the loop and suggested adding a max-turn break condition. Another common issue is over-prompting: asking the model to “think step by step” when a direct answer would suffice. Each unnecessary turn adds input and output tokens. Audit your workflows for turn counts and set hard limits to prevent runaway token consumption.

8. Share Token Data Across Workflows with Internal Anchors

Better visibility leads to better optimization. GitHub uses the token-usage artifact as a shared resource across multiple workflows. For example, the auditor workflow links directly to the optimizer's issue (item 4) via internal anchor tags in the report. This creates a web of accountability: developers can click through from a spike in token usage to the specific optimization suggestions. You can implement similar cross-linking in your own CI/CD pipeline by generating HTML reports with anchor links. This turns a raw log file into an actionable dashboard.

9. Automate the Optimization Loop

The auditor and optimizer are just the beginning. GitHub has started experimenting with a third workflow that automatically applies low-risk optimizations—like model downgrades or cache directives—after a review period. The idea is to close the loop from detection to resolution without human intervention. Of course, any auto-fix must be tested in a sandbox first. Because all workflows are defined in YAML and run in GitHub Actions, you can easily create a staging environment to validate changes before rolling them to production. The result is a self-healing system that continuously reduces token waste.

10. Monitor, Measure, and Iterate

Token efficiency is not a one-time fix; it's an ongoing practice. GitHub's daily audits have shown that what worked last month may become wasteful this month as models and usage patterns evolve. The key is to keep monitoring the token-usage.jsonl artifacts, watch for regressions, and run the optimizer regularly. Share your findings with the team and celebrate wins (like a 50% reduction in token cost for a heavily used workflow). By making token efficiency a visible metric—and automating its improvement—you ensure your agentic workflows remain affordable as they scale.

In the end, optimizing token usage is about being deliberate. The same automation that makes agentic workflows so powerful also makes them prone to silent waste. But with the right logging, auditing, and automated optimization loops, you can turn that weakness into a strength. GitHub's journey from April 2026 shows that the effort is worthwhile: we've seen dramatic cost reductions without sacrificing the hygiene and quality improvements that agentic workflows bring. Start small—maybe just implement token logging first—and build from there. Your future self (and your budget) will thank you.

Tags: