@jasondeanlee There's a nice paper I might be able to dig up about how asking agents to reflect on trajectories and update with learnings yields *worse* performance over time
Just because models can update their own prompts doesn't mean they're good at doing so
Jason Lee@jasondeanlee
I am convinced I am still useful. I spent about 10 days trying to write skill files and harness in codex with gpt 5.5 pro to replace myself : query gpt 5. 5 pro, sanity check, and ask it to rewrite or clarify or suggest vague new strategies. The results are not very good. I am still better at prompting than the harness
5:11 PM · Jun 14, 2026 · 913 Views