Theo Browne experiments with GPT-5.6 Sol inside Claude Code, arguing that the surrounding agent harness can materially affect design output, orchestration and token use. He contrasts dynamically written workflows with open-ended delegation and examines how detailed system instructions shape behavior.
The review includes examples of design differences, prescriptive interface guidance and formatting problems across harnesses. Theo also corrects an initial criticism after finding that an agent had published a requested link successfully. The conclusions are personal observations from an experiment, not a universal benchmark or a recommendation to disregard platform restrictions.
Watch on YouTube
