Takeaways from Anthropic's harness post

Some takeaways from Anthropic's post on agent harnesses:

1. Compaction can leak context anxiety. You get better results handing context off to a completely new agent.
2. Always use a separate agent to review an agent's work, since agents tend to think their own work is great.
3. A frontend design skill.md with 4 real design criteria makes a big difference to how impressive the app looks.
4. A 1M context window means your plan, code and review loop can bite off much larger chunks.
5. Orchestration is where it's going to be at for the foreseeable future, to get more results than a single agent can.
6. The challenge in creative tasks is for the agent to know if what it's built appeals to a human, but you can build tools that evaluate this to some extent.

That last one has been on my mind recently. Here are some of the more interesting quotes: