SI Agent News
DevelopingClaudeOpenAIGeminiAgents

LetThemBuild has AI coding agents check each other's work and show receipts

A new Show HN app has one AI agent write code while others review it and log the checks they ran. It is an early, unproven idea worth understanding.

1 pointShow HN score when we scanned the post

The short version

Why it matters

Many people now ask one AI to write code and then trust its own claim that it works. LetThemBuild tests a different habit: make a second and third agent prove their checks by showing what they ran. Even if this product goes nowhere, asking for evidence from an AI reviewer is a habit you can adopt today.

What LetThemBuild is

LetThemBuild is a desktop app, posted to Show HN (Hacker News's section for people to show off their projects), that coordinates several AI coding agents on one task. As letthembuild.com describes it, one agent leads and does the building, while three others review the work, run checks and report back.

The title names Claude Code, Codex, Gemini and OpenRouter. The product page also mentions Claude, ChatGPT, Gemini and Grok working out of the box. You can add hundreds more models through an endpoint you supply yourself. OpenRouter is a service that gives access to many models through one connection.

How the review room works

The maker calls the group a room. The lead does the work, the reviewers check it before it counts as done, and a warden looks at every action taken. The page does not say how the warden decides what to flag.

The most distinctive part is the receipt. Under every answer, the app lists the commands each reviewer ran and their exit codes. An exit code is the number a command returns when it finishes, and 0 conventionally means it succeeded. You can see what was checked, what passed and what failed.

The page says a reviewer that ran nothing is labeled as having run nothing. In other words, a vague approval gets no credit.

You also choose how hard the group works:

Leave it on Default and the room picks a level based on your request and what the lead actually did.

Where it runs and whose keys it uses

The group works inside a git worktree, which is a separate working copy of your project folder tied to the same code history. That keeps the experiment apart from your main files. It runs on macOS, Windows and Linux.

Nothing leaves your machine unless you pick the app's own keys, and the page says the app warns you before that happens. An API key is a password that lets software call an AI model. You can use the app's keys, bring your own, or run the tools through subscriptions you already pay for: Claude Code, Codex, Gemini CLI and Grok Build. Any endpoint compatible with OpenAI's format also works.

When a run finishes, Pull Back turns the work into its own branch and Open PR opens a pull request, which is a proposal to merge changes into a project. Move Back to This Mac carries on locally.

You can also hand a task to the cloud from the Mac, the web, an iPhone or a terminal command. From any device you can follow each step, approve a command or stop the run. A companion app, LetThemChat, shares the same account and sessions. It is for reading what the agents did, asking them questions and steering the next step.

How much to trust it

Not much yet, and the post's own record says so. When we scanned it, the Show HN had 1 point and 0 comments. That means almost no outside reader had tried it or commented. Everything above comes from the maker's own page.

The page we reviewed gives no benchmarks and no pricing. It points to documentation for how the room reaches agreement and how the money works, but we have not seen that material. Independent testing is still open.

The receipt idea is easy to like, but a receipt only proves a command ran. It does not prove the command was the right test. A reviewer that runs a shallow check and passes it still produces a clean-looking record. Someone still has to read what was run.

The idea worth borrowing

You do not need this app to use the principle. When an AI says its work is done, ask which checks it ran and what they returned. Treat a claim with no evidence as unreviewed.

The app also reflects a wider trend: the model underneath is increasingly swappable, and the useful work is in how several of them are organized. More reviewers means more output to read, which costs time and, if you use the app's keys, possibly money.

What to watch

Try it

Next time an AI coding assistant says a change works, ask it to list the exact commands it ran and what each returned, then rerun one yourself.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by letthembuild.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news