GitHub session limits matter because Copilot CLI and SDK runs can now stop at an AI credit cap that covers model calls, subagents, and background work instead of drifting into unbounded spend.
The expensive agent session is usually not the one you are watching. It is the one that keeps going after you assume the task is basically under control.
That is why GitHub's new session-limit control for Copilot CLI and SDK is more important than it first sounds. On July 1, GitHub announced that you can now set AI credit session limits so one active session has a defined spend boundary. In an interactive session, you can use /limits. In a noninteractive run, you can pass --max-ai-credits. GitHub says the counter tracks the whole session, including model calls, subagents, and background work like compaction.
This is a small feature on paper. Operationally, it is a big deal.
Old software budgets were often attached to seats, servers, or predictable usage bands. Agent work is messier than that.
A single task can branch. It can spawn helpers. It can reread context, call multiple models, or spend extra cycles cleaning up after itself. The cost profile is not just about whether an employee clicked "run." It is about how much work the system ends up deciding to do before it considers the task done.
That gets especially uncomfortable in unattended workflows. Maybe you kick off a batch review job, a migration helper, or a documentation cleanup lane before dinner. If the session keeps wandering, retries too much, or takes a slower path than you expected, cost can drift long before anyone notices.
GitHub's own launch copy hints at the core use case: automation where nobody is actively monitoring the agent's work.
Plenty of platforms can tell you after the fact that you spent money. That is not the same as giving the operator a practical stop-loss.
A monthly budget is coarse. A per-seat allowance is coarse. Even org-wide spend reporting is coarse when the real risk comes from one session that ran longer or deeper than intended.
Session limits change the control point. Instead of asking finance or admin tooling to explain the bill later, the operator can say up front: this session gets this much room, and then it stops.
That matters because it matches how agent work actually feels. The risky unit is often not the whole account. It is one active run.
GitHub says the limit covers model calls, subagents, and background work like compaction. That is the part that makes the feature credible.
If the cap only applied to the top-level prompt, it would not solve much. Modern agent sessions spill into helper tasks and behind-the-scenes steps. A real stop-loss has to follow the entire session footprint, not just the visible command the user launched.
GitHub also says the cap is soft because usage is only known after a response returns. That is an important caveat, and Butler readers should take it seriously. The point is not perfect accounting at the single-credit level. The point is that the session now has a bounded operating envelope instead of infinite implied permission.
The best early use is not to put tiny limits on everything and then wonder why jobs stop too early. It is to identify the unattended lanes that are already uncomfortable from a cost perspective.
Good first candidates include:
In those cases, a session limit is not just a spending tactic. It is also a behavioral nudge. It forces teams to think about what "enough work" means for a run before they hand autonomy to the tool.
That is healthier than pretending every agent task deserves an open-ended budget until proven otherwise.
Butler has been tracking a steady pattern across agent platforms: the product is not only the model anymore. The product is the control surface around the model.
Pricing stories, BYOK policies, review-depth settings, and now session stop-loss controls all point the same direction. As agents get more autonomy, platform buyers stop asking only "how capable is it?" They start asking "what keeps this inside the bounds I intended?"
Session limits are a clean example of that shift. They are not glamorous, but they are exactly the sort of feature that makes agent experimentation survivable in real teams.
GitHub's session limits matter because they give unattended Copilot work a native stop-loss. /limits and --max-ai-credits are really workflow-governance features wearing a billing label.
That is a healthy sign for the market. The agent platforms that win will not just offer more autonomy. They will offer better ways to bound it.
This article was researched and drafted with AI assistance, then reviewed and edited for clarity, accuracy, and editorial quality.