Two harnesses, two very different answers to the same question: what is actually stopping the agent?

The Approval Prompt Is Not the Sandbox

Codex ships 43,591 lines of sandboxing across four platform backends. Pi ships zero and says so in its README. Both are defensible positions. The one in between is not, and it is where most people are: approvals on, sandbox off. An approval prompt shows you a command string. It does not show you what that command reaches.

Three Agent Harnesses, One Loop

Claude Code's agent loop is 1,729 lines. Codex's is 983. Pi's is 794. Three teams, three languages, no shared code, and the loop lands in the same place every time. What does not converge is where each one draws its boundary, and that is the part you cannot change later.

Separately, nine kernel debugging pieces went up over the weekend. They are written versions of things that until now existed only as video, which meant they were not searchable, not skimmable, and not quotable. Bootloaders, GDB against real mode, eBPF hooking internals, memory reordering. Start anywhere:

• kprobe vs fentry, and what production observability misses
• Store to load reordering, x86 against ARM64
• Real mode to protected mode, building GDT and IDT by hand

The other six are on the blog.

Harrison