The long session myth: why 10 minutes is enough

267 coding agent sessions tracked since December 2025 on a Symfony/TypeScript project. The duration distribution:

< 10 min     165 sessions (62%)
10 - 60 min   61 sessions (23%)
1 - 4 h       21 sessions  (8%)
> 4 h          20 sessions  (7%)

Most agentic work happens in micro-sessions. Not 2-hour marathons with the agent — short, focused interactions, then back to the code.

What happens in a 10-minute session

One precise question, one answer, maybe one adjustment. Typical examples:

“Why does this PHPUnit test fail with a 500 on this endpoint?” — the agent reads the test, the controller, identifies the missing mock. 3 minutes.

“Generate fixtures for this entity with its relationships.” — the agent produces the file, you check, adjust one field. 5 minutes.

“Refactor this switch to a strategy pattern following project conventions.” — the agent reads the existing code, the AGENTS.md conventions, produces the refactoring. 8 minutes.

These are complete deliverables. Not unfinished sessions.

Why long sessions are an anti-pattern

Beyond 30-40 exchanges, context rot degrades response quality. Early instructions lose their influence. The agent starts contradicting its own earlier decisions.

The 20 sessions over 4 hours in the data are almost all “catch-all” sessions — chaining unrelated tasks, drifting debugging, exploration without clear goals. The outcome/time ratio on these sessions is significantly lower.

Context rot isn’t the only cause. On open models a second ceiling stacks on top: I replayed the same trivial task three times on the opencode Go models, and roughly one run in three comes out wrong. The verdict from practitioners who pushed every one of them is that they collapse beyond 5-10 minutes of work. Context degradation on one side, a capability ceiling on the other: two distinct mechanisms landing on the same instruction.

The reflex to build

The right question isn’t “how to make sessions last longer” but “can I break this into a 10-minute session?”

If the answer is no — if the task genuinely requires broad context and sustained reasoning — then a long session is justified. That’s 15% of cases. For the remaining 85%, the micro-session is more effective, more predictable, and less prone to context rot.

The context window isn’t a tank to fill. It’s a resource to use sparingly.

I show these numbers systematically to a dev who’s new to agents — it defuses the assumption that a good session has to be long, before that assumption even sets in.

Guide: Managing context in multi-agent workflows Guide: Agentic workflow principles