‹ BackHN Continuity

Thread

Unreal Agent

246 points · 125 comments · trollied

  1. tekacs · · focus · HN ↗
    The headline graph is kind of bizarre.

    For some reason they're comparing their harness running on Astra xhigh to Codex with Astra max?

    ---

    Also worth noting that OpenAI just added support for async tool calling to their harness, which isn't 1:1 with this approach, but is slowly ramping up in being able to provide something similar.

    A big part of why Codex uses so many tokens is that it basically hot loops on polling tasks it starts for... absolutely no good reason: <a href="https:&#x2F;&#x2F;www.reddit.com&#x2F;r&#x2F;codex&#x2F;comments&#x2F;1wdlp7q&#x2F;weve_discovered_the_issue_behind_codex_harness&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.reddit.com&#x2F;r&#x2F;codex&#x2F;comments&#x2F;1wdlp7q&#x2F;weve_discove...

    I fixed it on my fork of Codex too, also back in Jan&#x2F;Feb – I keep this patch rebased, for anyone who wants it: <a href="https:&#x2F;&#x2F;github.com&#x2F;tekacs&#x2F;codex&#x2F;commit&#x2F;9ffcf8db9078eae43d4111ff94259795c1e962c9" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;tekacs&#x2F;codex&#x2F;commit&#x2F;9ffcf8db9078eae43d411...

    It results in token savings similar in scale to those displayed here by Unreal.

    ---

    My harness has used a slightly fancier version of the approach that Unreal is using since ~Feb, and... it definitely works excellently, but it&#x27;s also assuredly smoother with Astra and other recent models that are more aware of async tool calling.

    1. NostraDavid · · focus · HN ↗
      To reduce the codex polling I slapped this into my ~&#x2F;.codex&#x2F;config.toml:

        [features.multi_agent_v2]
  enabled = true
  wait_agent_enabled = true
  min_wait_timeout_ms = 10000 # 10s
  default_wait_timeout_ms = 300000 # 5m
  max_wait_timeout_ms = 3600000 # 1h
      Seems to do the job and reduce usage; I just ran Astra for ~5 hours (using a goal) and it used the last 30% of my usage. And now they released GPT-6 Sol and Luna (which is basically 5.6 Sol and Luna, but a bit better and also 50% cheaper) ;_;
      1. AmazingTurtle · · focus · HN ↗
        Perfect, I will incorporate this as default as well as the commit from the other guy into my own codex fork <a href="https:&#x2F;&#x2F;github.com&#x2F;AmazingTurtle&#x2F;codex" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;AmazingTurtle&#x2F;codex btw. I&#x27;m rebasing on 0.156.0 right now
Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.