I notice I get more success if I monitor the success or failure of the composer agent and just reset to the prompt state checkpoint and let it try fresh than if I just let it keep building on its efforts.
If othersfind the same paradigm to be effective, a feature letting the agent detect its attempt to achieve the prompted outcome has failed and reset automatically to the prompted checkpoint to just try again (as if I manually just re-ran the initial prompt & reset at the checkpoint) could be a really nice step forward.