GitHub cut the token bill by giving AI agents fewer tools to wander through

In agentic systems, cost often grows not because the model is solving a harder problem, but because the environment keeps offering too many options. If an agent sees a broad set of tools, resources and schemas through the Model Context Protocol, every call can drag extra metadata into the context window. GitHub therefore pruned unused MCP tools and moved some operations back to the GitHub CLI, where a command can be more direct and cheaper than routing the same job through a general agent tool layer.

That distinction matters. MCP is useful when an agent needs structured access to systems, but not every task improves just because it passes through an agent protocol. For routine GitHub operations, gh can be the shorter path: fewer tool descriptions, less indirect context, and less room for the model to spend tokens discovering what a script already knows.

The second part is measurement. GitHub introduced a token-usage.jsonl artifact and an Effective Tokens metric to track spend across models and spot regressions. That brings agentic workflows closer to mature DevOps practice: if a build gets slower, teams respond; if a test starts failing, teams respond; if an agent suddenly spends more tokens for the same job, that should also be a signal, not a footnote on an API bill.

The daily “auditor” and “optimizer” agents are not decorative automation in that setup. Their job is to inspect where tokens are being spent, which tools are no longer needed, and where a workflow can be narrowed without losing function. In effect, agents are being used to supervise other agents, but with a concrete metric and an execution artifact that remains after the run.

For teams building CI agents, the lesson is blunt. It is not enough to add a model, connect MCP servers, and assume intelligence will optimize cost by itself. Teams need a tool inventory, daily review, a clear metric, and regression thresholds. GitHub’s example also shows that optimization does not have to mean a weaker model or a cruder prompt. Sometimes the largest gain comes from no longer giving the agent too many doors to walk through.

In the wider context of GitHub Copilot and increasingly agentic development workflows, this is the practical line between demo and production. Agentic CI can be useful, but only if it is visible, measurable, and narrow enough that cost does not become hidden infrastructure growing faster than the value it creates.