2 ms·

Oh-my-pi – The Harness Playbook

35 points by maherbeg 2mo ago

wolfiesch 2mo ago
BILLIONS MUST OMP
Centigonal 2mo ago
The more people use agents, the more they write like agents. The author seems to have put significant effort into this post, but when I read something like the below, my brain just immediately checks out. > # What one authority buys > With state and transcript in one tree, several hard problems reduce to the same operation. > Rewind is a DOM diff. Diff the current materialization against the target state. A <subagent> element disappeared? Terminate it by destroying the element. One appeared? Resume or spawn it by creating the element. The delta itself is the complete lifecycle work list. A sentence like "The delta itself is the complete lifecycle work list" gives me mental indigestion. I have to spend 30 seconds dissecting it to figure out what the author is trying to say. When every other sentence is like this, the post becomes unreadable.
- jdkoeck 2mo ago
  Did the author really put significant effort here? The LLM speak suggests the opposite.
  - solarkraft 2mo ago
    > These are not market personas. They are architecture tests. Indeed.
- smcleod 2mo ago
  It all smells pretty load-bearing to me.
  - cheikhcheikh 2mo ago
    But is it byte for byte identical ?
    - smcleod 2mo ago
      You're right to push back.
poppafuze 2mo ago
Slopmaxxing text.
localroast 2mo ago
Oh-my-pi is designed to be slim apparently but has a ton of useless features with little empirical justification. Pi is a good harness- probably the best, but OMP not so much. One thing I hate about OMP more than anything else though: hash anchored edits. they don't work in practice; most models I've tried have spent more turns on repairing their broken edits than on the actual task.
- riedel 2mo ago
  I like it a bit for the inspiration what could make sense, however, as there is little ablation studies, it is really difficult to understand the effect of most of the features.
- coffeeindex 2mo ago
  From what I’ve read/heard, a lot of the power of Pi is its extensibility but it’s a hard sell for me to try it out just because I can extend it. How is it out of the box?
  - localroast 2mo ago
    You'd probably want at least some packages, like @henryqw/pi-auto-compact The system prompt is very slim, but Pi is obviously barebones. Biggest problem I find is there's no real configuring of when to compact, compaction is going to be manual. About half the packages available are various solutions to this problem. Out of the box it doesn't appear to support parallel tool calls, or at least the model I normally use (GLM 5.3 Flash) doesn't seem to make use of it. maybe just needs a simple note that parallel tool calls are available in the system prompt? I observe parallel tool calls with OMP. There's no MCP support either, but of course there's an extension for that as well. I'd be careful when using different providers or your OpenAI subscription. I'm not sure if Pi includes reasoning chains in the running conversation with each API call for every provider. I'm pretty sure reasoning is kept across the session with OpenRouter, at least. I always start Pi with '--no-skills'. I find it strange how they include skills as part of the base package... I'd be interested to know how well it works with an OpenAI or ZAI subscription.
8MEM 2mo ago
[flagged]

35 points by maherbeg 2mo ago

BILLIONS MUST OMP

The more people use agents, the more they write like agents. The author seems to have put significant effort into this post, but when I read something like the below, my brain just immediately checks out. > # What one authority buys > With state and transcript in one tree, several hard problems reduce to the same operation. > Rewind is a DOM diff. Diff the current materialization against the target state. A <subagent> element disappeared? Terminate it by destroying the element. One appeared? Resume or spawn it by creating the element. The delta itself is the complete lifecycle work list. A sentence like "The delta itself is the complete lifecycle work list" gives me mental indigestion. I have to spend 30 seconds dissecting it to figure out what the author is trying to say. When every other sentence is like this, the post becomes unreadable.

jdkoeck 2mo ago
Did the author really put significant effort here? The LLM speak suggests the opposite.
- solarkraft 2mo ago
  > These are not market personas. They are architecture tests. Indeed.
smcleod 2mo ago
It all smells pretty load-bearing to me.
- cheikhcheikh 2mo ago
  But is it byte for byte identical ?
  - smcleod 2mo ago
    You're right to push back.

poppafuze 2mo ago

Slopmaxxing text.

localroast 2mo ago

Oh-my-pi is designed to be slim apparently but has a ton of useless features with little empirical justification. Pi is a good harness- probably the best, but OMP not so much. One thing I hate about OMP more than anything else though: hash anchored edits. they don't work in practice; most models I've tried have spent more turns on repairing their broken edits than on the actual task.

riedel 2mo ago
I like it a bit for the inspiration what could make sense, however, as there is little ablation studies, it is really difficult to understand the effect of most of the features.
coffeeindex 2mo ago
From what I’ve read/heard, a lot of the power of Pi is its extensibility but it’s a hard sell for me to try it out just because I can extend it. How is it out of the box?
- localroast 2mo ago
  You'd probably want at least some packages, like @henryqw/pi-auto-compact The system prompt is very slim, but Pi is obviously barebones. Biggest problem I find is there's no real configuring of when to compact, compaction is going to be manual. About half the packages available are various solutions to this problem. Out of the box it doesn't appear to support parallel tool calls, or at least the model I normally use (GLM 5.3 Flash) doesn't seem to make use of it. maybe just needs a simple note that parallel tool calls are available in the system prompt? I observe parallel tool calls with OMP. There's no MCP support either, but of course there's an extension for that as well. I'd be careful when using different providers or your OpenAI subscription. I'm not sure if Pi includes reasoning chains in the running conversation with each API call for every provider. I'm pretty sure reasoning is kept across the session with OpenRouter, at least. I always start Pi with '--no-skills'. I find it strange how they include skills as part of the base package... I'd be interested to know how well it works with an OpenAI or ZAI subscription.

8MEM 2mo ago

[flagged]