Inside the Harness: Four Months of Being Wrong About pi

On July 20th, a coworker asked this in our #pi channel [source: Slack thread, 2026-07-20 19:03 CDT]:

Why did @nicknisi say he no longer thinks pi is a good idea before I get any deeper into it?

My reply came in two messages [source: same thread, ts 1784592942 and 1784592964]:

Haha I merely said that there’s real innovation in harnesses that shouldn’t be ignored!

Love pi though! :homer-disappear:

Zack: “WHAT DO YOU MEAN MAN” [source: same thread]

Fair. This is the answer I owed him, rebuilt from the record.

Day one and day two, from the logs

At 2:36pm on Wednesday, March 18th, I posted in #pi [source: Slack, 2026-03-18 14:36 CDT]:

lol, love that this is also Pi’s domain: https://shittycodingagent.ai/

At 3:07pm I asked a fresh pi session “what can pi do” [source: session log 2026-03-18T20:07Z]. By 3:27pm I’d pasted my Claude Code statusline into a dotfiles session and asked pi to recreate it [source: session log 2026-03-18T20:27Z]. The statusline extension, a tmux status bridge, a Night Owl theme, and spinner verbs all landed as commits that evening and the next morning [source: dotfiles 085af4a, aa78659, fbcc73b, 4100966].

Day two, I posted Lydia Hallie’s tweet about Claude Code running shell commands inside skills, then asked pi: “Can you execute commands in skills? Read this tweet to see what I mean.” My entire follow-up prompt: “do it.” By 9:04am I was back in the thread: “I asked pi if pi could do this too and it made an extension 🤯” with a link to what’s now dynamic-skills.ts in the monorepo [source: session log 2026-03-19T13:47Z; Slack, 2026-03-19 09:04 CDT].

Five months of usage numbers

Pi commits in my dotfiles: 19 in March, then 2, 5, 1, 4 through July [source: git log on home/.pi paths, by month].

Part of why: sometimes pi gets stuck on things, especially if I’m closing my laptop/switching networks, etc. This has cost me more than once hours where I think it’s working but it’s actually doing nothing.

The one that stung: I was building a weekend project, hoping to get it presented at work the following Monday. It worked all weekend, or so I thought. In the end, I was able to present it, but it wasn’t as robust as I had hoped.

I told people it wasn’t worth the switch. The stuck sessions were part of it. The bigger part: my whole stack (case, Fleet, the tmux scrapers) only spoke Claude Code.

What changed in August

On August 1st I opened a pi session with this prompt [source: published post blockquote]:

so far I’ve been using pi like i’ve been using claude code. I know I can go deeper and make things better with pi, but I’m not sure exactly how. What are the capabilities beyond a simple claude code-like coding harness?

On August 3rd I asked #pi: “Anyone have their pi configs open source? I’d love to scour them!” and shared mine [source: Slack, 2026-08-03 11:07 CDT]. In that thread I also wrote: “my perception is I should be thinking about subagents differently in pi and I’m still thinking about them from a claude perspective” [source: same thread].

August 6th

At 10:36am I opened the pi session that did the move [source: session 2026-08-06T15:36Z]. The commit landed at 1:23pm: 53 files changed, 25 insertions, 25,635 deletions [source: dotfiles 0e39074]. Along the way, a research pass caught my trust.json sitting in the public repo [source: dotfiles 94fcef9].

What replaced all of it lives in one extensions monorepo: 26 packages at publish time [source: pi-extensions ccff6ab]. The statusline that started everything is 519 lines of TypeScript [source: same commit]. Fleet became a pi extension that asks the agent what it’s doing instead of scraping tmux for it.

The next ten days: 188 pi sessions, 17 Claude Code sessions [source: my count at publish; method-dependent].

The answer I owed Zack

So when Zack asked, the honest answer was this: I was just skeptical that issues like this were my fault and it would take an investment in making sure the tools work, which ended up being the case. Still, at the time I wasn’t ready to make that investment and I just wanted to get things done.

I think that’s a big dichotomy when you’re working on these tools. You need to show that they work, but you also need to constantly improve and make your workflows better, and you’re constantly thinking about both of them.

Was rebuilding everything on one niche harness’s API worth it? Because I learned to go beyond what the big AI companies are doing. There’s so much innovation at the harness level and I have learned so much. That alone is worth it, even if it goes away tomorrow.

Do I miss anything from the 25,635 deleted lines? Not at all. I feel more productive than ever and I’m having more fun than ever.