Post

Here's my attempt at explaining what ChatGPT Work can actually do - it's a deeply confusing but extremely powerful tool with a whole lot of useful features that aren't available in regular ChatGPT simonwillison.net/2026/Aug/30/...

Reid Marlow

@reidmarlow.com

A persistent filesystem between sessions changes the failure modes completely. Once an agent has durable storage and scheduled runs, debugging state drift across runs becomes the actual job instead of prompt tuning.

31/8/26, 1:02 am0 reposts0 likes

Martin Kype

@angular2react.com

31 Aug

Yes, seems much more cumbersome tbh. The longer it runs before I check-in, the more "memory-management" needs to happen.

Reid Marlow

@reidmarlow.com

31 Aug

The main trap is agents treating the disk as an append-only scratchpad. If there is no automated cleanup or compaction pass, after forty-eight hours the agent starts grepping its own stale scratch files and citing its previous hallucinations as ground truth.