I let an agent run for four hours on one task. It ended up re-suggesting the same approach it had already tried and abandoned two hours earlier. The conversation log even said so. It just couldn't see it anymore.
That's the failure mode nobody warns you about. Sessions don't get smarter as they get longer. They rot.
What rots, specifically
Early in the session the agent has sharp context: what you tried, why it failed, the decision you locked in. A few hours in, that's a blur of tool calls and edits. It starts contradicting its own earlier choices, re-exploring paths you already closed, and treating a settled decision as open again.
It's not the model getting dumber. It's that the relevant facts are now buried under hundreds of messages, and the ones that matter least are the ones closest to the top of what it attends to.
The trap is sunk cost
The longer the session runs, the harder it feels to throw away. "But it already knows the codebase." That feeling is wrong. The thing it learned is cheap to recreate; the thing it's now forgetting is expensive to rediscover.
A fresh session looks slower for ten minutes. It's faster for the whole rest of the task.
The habit that works
Every hour or so, I stop and write one short markdown checkpoint: what changed, what's decided, what was ruled out and why. Then I start a new session and paste that checkpoint as the opening context. The agent gets the decisions without the baggage of a thousand messages.
For work involving API decisions this is the natural place to be explicit. The endpoint shape we agreed on, the contract we're keeping — that should live in the spec file, not only in chat memory. A fresh session reads the spec and is instantly caught up.
The rough rule I use
If the agent suggests something we already rejected, or re-explains a file it edited an hour ago, the session has rotted. Don't prompt harder. Checkpoint and restart.
How long is your longest successful agent session before you notice it going in circles?
Top comments (0)