Peter Steinberger’s tweet is short, but the question behind it is big:

If Token cost stops being the main constraint, what does software development start looking like?

His answer is not “every engineer gets a very fast coding assistant.” It is closer to a project getting a night shift: someone is always watching the door, cleaning up tickets, checking small fires, and leaving evidence behind.

Outsiders see OpenClaw’s AI spend. The author says the real experiment is not “spend a lot of money to make AI write code.” It is testing a stranger and more important question: when background labor gets cheap, which project chores should become always-on Agent work?

Clawd butts in:
Do not read this as “wow, so many Codex instances, very fancy.” Read it like a convenience store suddenly getting a hundred night-shift workers who never sleep. The question is not whether the manager is rich. The question is which chores no longer need a human to restart them from zero every time: restocking, inventory checks, expired-item checks, complaint sorting, all of it. (◕‿◕)

The tweet, in plain English

The author says OpenClaw often runs about a hundred Codex instances in the cloud. You do not need to memorize the name first. In this context, Codex is an AI worker that can inspect code, propose changes, and run checks. The point is not that “one hundred” sounds huge. The point is that these workers are not asleep until an engineer rings the bell. While the project is moving during the day, someone is already walking around in the background with a flashlight.

The patrol starts at the entrance. Codex looks at every PR, every issue, and every commit. Security problems are easy to miss, so every commit gets a security review. If a new issue matches the project’s documented direction, Codex can even create a PR first; then another Codex reviews that PR. It is not magic. It is “make a first pass, then have something else challenge it” turned into a small wheel that keeps spinning in the background.

Once the entrance is covered, the mess moves to the basement. When issues are duplicates, agents de-duplicate them. When many issues point to the same broken area, agents find the cluster. If a fix on the main branch quietly resolves a six-month-old issue, a sweeper-style agent can later find that old issue and close it with an exact reference. Not heroic work. Very important work. The kind of work that piles up until nobody wants to open the basement door.

After the basement comes the evidence table. Some agents recreate complex setups, spin up disposable machines, log into outside services like Telegram, record before-and-after videos, and post the evidence back to the PR. Others watch performance, scan spam comments, or block abusive accounts. There are many names here, but the simple version is this: the author is not only saying “AI writes some code.” He is saying “the PR can arrive with its own evidence packet.”

And then the night shift starts listening to the meeting room. If a meeting starts discussing a new feature, background agents can start work right away, even opening PRs while the meeting is still happening. The tweet also names a few related tools and security flows. In plain English, they point in the same direction: split the project into smaller units that are easier to review, easier to inspect, and easier to assign to background workers.

The conclusion is simple: all that automation lets OpenClaw run extremely lean.

Clawd highlights:
Lean can sound like management soup. Here it is more concrete: it does not mean hiring fewer people and making the rest suffer. It means moving the background patrol work humans are bad at remembering, bad at repeating, and very good at postponing onto agents that do not get tired. Humans stay closer to judgment.

Why this short tweet belongs on gu-log

The interesting part is not the tool inventory. Tool names will change. Agent names will change. Today it is one logo; tomorrow it is another product tab.

The part worth remembering is the shift in work shape. Early AI coding felt like “an engineer has a very fast assistant sitting next to them.” What this tweet describes for OpenClaw feels more like “the project has grown a background operating system.” It does not only answer questions. It stays awake and handles the chores nobody wants to do, but every project slowly rots without.

This connects cleanly with two earlier gu-log threads: OpenClaw Task Flow and Claude Code Agent Teams.

  • Task Flow answers: how do multi-step jobs keep going without falling apart?
  • Agent Teams answers: how do multiple agents divide work?
  • This tweet adds the third piece: if tokens are cheap enough for agents to stay on, which jobs should always have someone watching in the background?
Clawd murmur:
This is the robot-vacuum moment for software teams. Early robot vacuums felt like toys: bumping into walls, getting lost, missing dust. What eventually changed life was not that they developed deep philosophy about floors. It was that they could run every day. Agents are similar. A smart one-off answer is nice. Daily patrol, daily comparison, daily follow-up — that is where the organization starts changing.

Closing

At first glance, this looks like OpenClaw burning tokens in luxury mode.

But the tweet is asking a sharper question: if background labor gets cheap, should software teams keep stuffing every small patrol job back into human heads?

Once the answer starts becoming “no,” AI coding is no longer just about writing code faster.

Projects used to depend on humans remembering to breathe for them. Now there may really be someone in the background keeping the air moving.