I believe Anthropic will iterate their software (e.g. Claude Code) to the point where you don't need to apply anything from your side to get decent results. For example, no one asks currently Claude Code to do "Chain of thoughts". It just works. Probably in spaces where you can have checkable outputs (e.g. programming, software), you don't need to tweak anything and learn about it. However, there are spaces which are kind of art (e.g. writing) where you need to tweak to get some decent results. That said, you will not get great results anyway
Your framing that the playbook explains almost none of the 8x is the important line, and I think the missing variable is who reviews the output. Anthropic merges more per day because the people reading the diff are the same people who built the model's habits, so verification is fast and trusted. Outside that, Faros AI found across 22,000 developers that throughput per developer rose 33.7% while median review time rose 441.5%, and LinearB found AI-assisted PRs merge within 30 days 32.7% of the time against 84.5% unassisted. Same context engineering, very different merge rate. Code output is the cheap half now. The scarce half is someone qualified to say yes.
Thanks for writing the plain version. Compaction over replaying is the right default, though compaction is lossy in a way replaying is not, so which file survives matters.
I believe Anthropic will iterate their software (e.g. Claude Code) to the point where you don't need to apply anything from your side to get decent results. For example, no one asks currently Claude Code to do "Chain of thoughts". It just works. Probably in spaces where you can have checkable outputs (e.g. programming, software), you don't need to tweak anything and learn about it. However, there are spaces which are kind of art (e.g. writing) where you need to tweak to get some decent results. That said, you will not get great results anyway
Some thoughts on this lately here https://www.thetrueengineer.com/p/i-tested-every-ai-model-the-same
Your framing that the playbook explains almost none of the 8x is the important line, and I think the missing variable is who reviews the output. Anthropic merges more per day because the people reading the diff are the same people who built the model's habits, so verification is fast and trusted. Outside that, Faros AI found across 22,000 developers that throughput per developer rose 33.7% while median review time rose 441.5%, and LinearB found AI-assisted PRs merge within 30 days 32.7% of the time against 84.5% unassisted. Same context engineering, very different merge rate. Code output is the cheap half now. The scarce half is someone qualified to say yes.
Thanks for writing the plain version. Compaction over replaying is the right default, though compaction is lossy in a way replaying is not, so which file survives matters.
I turned these concepts into a runnable tutorial using Python, now with 600+ GitHub stars: https://github.com/hardness1020/awesome-agent-architecture/tree/main/sections/08-context-management