3 ms·

I'm still inclined to question whether the agent harness needs to be a million lines, or even a hundred thousand lines, in the first place. I mean, isn't one su

I'm still inclined to question whether the agent harness needs to be a million lines, or even a hundred thousand lines, in the first place. I mean, isn't one supposed to just hook up tool calling and let the LLM's intelligence do the rest? At least, that's the way it appears from things like tptacek's post about writing an agent. https://fly.io/blog/everyone-write-an-agent/ https://fly.io/blog/everyone-write-an-agent/ Then again, I'm also sympathetic to the idea that the anti-AI backlash has also provoked an unreasonable backlash against development speed and scale themselves.

seanmcdirmid 21d ago
Harnesses are a lot of prompting and sandboxing also. Then you need to add goldfish reminders when it makes mistakes depending on model quality. It isn’t a lot of code, it definitely isn’t a lot of code that you can’t build with an agent.
simonw 21d ago
Yeah, I'm with you on that one. It seems to be a recurring pattern though. The X.ai coding agent harness is 840,000+ lines of Rust - I dug into that here: https://simonwillison.net/2026/Jul/15/grok-build/ https://simonwillison.net/2026/Jul/15/grok-build/ Codex is 1.4m of Rust, and that's not including the extra code for the closed source desktop app. Meanwhile Pi is 265,000 lines of TypeScript (130,000 if you exclude tests), so even the agent with a reputation for being smaller is pretty huge.
tptacek 19d ago
To be fair, I was mostly pointing out that you can get to the kernel of a useful agent very quickly and without much code, not making a claim that coding agents would or should generally be tight. Competitive coding agents in particular are tricky, in part because the two mainstream agents (Codex and Claude) are big RL targets and you generally want to be in their slipstream --- or so I'm told. On the other hand: I have basically zero interest in writing a coding agent (there are so many effective ones already), but I write lots of little agents, for instance in my music player, and those are as code artifacts pretty trivial.

seanmcdirmid 21d ago

Harnesses are a lot of prompting and sandboxing also. Then you need to add goldfish reminders when it makes mistakes depending on model quality. It isn’t a lot of code, it definitely isn’t a lot of code that you can’t build with an agent.

simonw 21d ago

Yeah, I'm with you on that one. It seems to be a recurring pattern though. The X.ai coding agent harness is 840,000+ lines of Rust - I dug into that here: https://simonwillison.net/2026/Jul/15/grok-build/ https://simonwillison.net/2026/Jul/15/grok-build/ Codex is 1.4m of Rust, and that's not including the extra code for the closed source desktop app. Meanwhile Pi is 265,000 lines of TypeScript (130,000 if you exclude tests), so even the agent with a reputation for being smaller is pretty huge.

tptacek 19d ago

To be fair, I was mostly pointing out that you can get to the kernel of a useful agent very quickly and without much code, not making a claim that coding agents would or should generally be tight. Competitive coding agents in particular are tricky, in part because the two mainstream agents (Codex and Claude) are big RL targets and you generally want to be in their slipstream --- or so I'm told. On the other hand: I have basically zero interest in writing a coding agent (there are so many effective ones already), but I write lots of little agents, for instance in my music player, and those are as code artifacts pretty trivial.