Steipete Replicates Codex, the Save-Me-Money Prompt & Warden Burns $25K on Sonnet
Agentic Coding & Agent Harnesses
Steipete's Saturday: four agent tools in one day, all open source. Peter Steinberger spent May 23 shipping a connected stack:
/autoreviewrunning 5 hours on a subagents refactor and finding pre-existing bugs. The skill (openclaw/agent-skills/skills/autoreview) runs across the changeset (vs.clawpatch, which does a full-repo scan) and surfaces issues the original implementation already had — not just the diff. "5 hours of autoreview finding real issues is the signal. Humans skim. The agent reads every line," replied one viewer. Steipete's tip in the thread: "tell codex to maintain a scratch-log while it works on bigger refactors with decisions it had to make, tradeoffs, review fixes, so later on you can read through which tradeoffs the agent made, what you forgot to specify etc." 330 likes, 32K views. Thread./autotriagereadsVISION.mdand verifies fixes with computer vision on a VM. Issues/PRs that (a) fit the project's vision, (b) are inferrable in code with high confidence, (c) have a clear fix, and (d) can be live-tested are now worked on autonomously by codex. The verification loop usescrabbox.sh(a new Parallels backend) to drive a VM and read the resulting UI; Steipete reviews suggestions manually. To skip typing issue text into prompts he built an issue browser intorepo.barthat parses common clipboard formats. 630 likes, 42K views. Thread.Codex-in-the-cloud on Cloudflare firecracker boxes. "Codex replicated itself, basically." Compute-limited locally, Steipete wrapped codex in Cloudflare's firecracker sandboxes — running tests via
crabboxwhen the firecracker box isn't beefy enough — and serves it through Ghostty compiled to WebAssembly. Minutes later @cnakazawa posted Cloudsail — independently built, same idea,npm install -g cloudsailto spin up a per-task Cloudflare Sandbox with a shell, Codex and GitHub access (tokens never exposed to the sandbox). On Cursor agents as an alternative, Steipete: "Yeah but it's not open source and limited to their stuff." 553 likes, 46K views. Thread.release.barGitHub dashboard. Public per-user view of repos, open issues/PRs, last released version, and commits since last release: release.bar/steipete. Includes a Trust Score — "if you're a clanker or not :)" 690 likes, 42K views (and his single highest-engagement tweet of the day). Thread.
"Please save me money" — auto-mode agents will actually do it. Thariq's drive-by reminder that you can just ask the agent to find cost leaks hit 52K views overnight, mostly off the chain of follow-ups establishing that yes, he was running auto-mode, yes it required explicit approval before destroying anything, and yes it was old unused Redis from a previous startup. The genre this gestures at — not "build me an app", just "find where I'm quietly bleeding money and stop it" (paraphrasing @iisanidhya) — drew several variations of the "takes down production database" joke and one substantive critique from @closermethod: "a prompt is a thing you remember. a skill is a thing that runs without you. that's the whole difference." Thread.
AI Economics & Spend
Sentry's Warden burns $25K/month on Sonnet — and that's described as too cheap. David Cramer (CTO, Sentry) disclosed that Warden — their security-review agent that fires on every commit on every PR org-wide — is at $25K in costs this month using almost exclusively Sonnet, "and we haven't spent the time optimizing yet." RT'd by mitsuhiko with the framing "we're still a couple orders of magnitude off from where costs need to be for this level of capabilities. Or we need capabilities to jump several orders of magnitude (which seems less likely)." Cramer's justification was a straight cost-of-risk argument: "any vuln that passes through is unbounded risk + we pay bounties on them if they're ever found." Vaibhav Srivastav floated the harness-swap angle: "unironically, I do think they'd be at least ~40% cheaper if they used GPT-5.4 / 5.3-Codex w/ Codex harness." 75 likes, 40K views. Thread.
Research & Limits
swyx co-signs an adversarial-world-models read of transformer limits. Quoting a post that frames RL as "learning from interventions" (versus supervised learning on observations), swyx tied it back to the adversarial world models essay he and @ankit2119 wrote earlier this year: "throwing more params, more power, more everything at a demonstrably inefficient paradigm will be outclassed by the simple solution that can hypothesize and seek truth rather than backfit a house of cards — although the bitter lesson is it is simpler to scale and we may hit AGI anyway because human intelligence just isn't that smart nor plentiful." The reply thread converges on the same concrete failure mode: "agents still struggle with tasks that require real error anticipation. It's not retrieval or planning that breaks first, it's not having a model that can simulate what goes wrong" (@jahanzaibai). Thread.
Hot Takes
Theo on Jarred Sumner's blog-post procrastination: "Rewriting Bun in Rust: 6 days. Writing a blog post about it: 2+ weeks." The follow-up to last week's Bun-in-Rust drop turned into the agentic-coding meme of the day — 75K views in six hours. The most popular reply (Wayne Allan): "IMO LLMs are nowhere near as good at prose and creating a memorable narrative as they are at code." Honorable mention to amul.exe's hypothesis: "there is no test-harness for the blog use case. Verifiability is the first condition for LLMs to converge." Thread.
mitsuhiko: "AI lets you do insane things." Pointing at a pi#4922 PR — the kind of fix-shift+enter-detection-across-terminals work that historically nobody wanted to do — and the reactions are mostly other people producing the same kind of patch for their own setup. Jarred Sumner in the replies, on the FFI choice: "smart you didn't do NAPI c++ bindings for this, simpler this way. Though this should just be a runtime api this kind of hack shouldn't be needed." Thread.
Theo flags "branch off with another model" as the killer chat UX. "The 'branch off with another model' feature in T3 Chat is still my favorite thing ever. So easy to get a 2nd opinion." Forking a chat into another provider mid-conversation to compare answers — a feature an increasing number of agentic harnesses are now leaning into. Thread.