GitHub has published a detailed look at how it is reducing the cost of agentic workflows that run inside GitHub Actions, a useful signal for teams starting to put AI agents on every pull request.

What changed

According to GitHub, the first step was adding workflow-level observability. Its internal runs now emit a token-usage.jsonl artifact that records per-call token usage, cache activity, model, provider, and timestamps. GitHub then added two recurring workflows: one that audits recent usage and another that inspects expensive runs and proposes conservative fixes.

The optimizations themselves are fairly practical. GitHub says one common waste source was carrying large unused MCP tool schemas in every request. Another was using MCP for deterministic GitHub reads that could be handled more cheaply with pre-run gh CLI steps or proxied CLI calls. The company says those changes produced meaningful savings across several internal workflows, with double-digit improvements in its published examples.

Why it matters

The important takeaway is not just lower token spend. GitHub is treating agent costs as an engineering problem that can be measured, reviewed, and tuned like the rest of CI. It also says the same auditing and optimization workflows are available through its GitHub Agentic Workflows ecosystem, which makes this more than a one-off blog post.

For developers building agent-heavy CI, the conservative read is simple: before adding more agents, instrument the ones already running and remove work that never needed model reasoning in the first place.