The Build Hypothesis logo, with the tagline Build it. Break it. Prove it., set over artwork of a developer at a monitor, an AI face, a rocket launch, a server room and a collapsing bridge

An experiment in public

What can AI coding agents actually build?

Same challenge. Different agents. A fixed time budget. Each coding agent gets the same versioned spec and builds on its own. Then the results get tested, compared and shown with the evidence.

01 The format

Build it. Break it. Prove it.

Not "can AI write code?" but what an agent can deliver with a serious product spec, limited time and a fair chance to work on its own.

  1. 01

    Write the spec

    An ambitious app, game or tool, fully specified before any agent starts: behavior, constraints, deliverables and acceptance criteria. Versioned and published.

  2. 02

    Build it

    Claude Code, Codex and Cursor CLI each get the same spec, environment and time budget. They build independently in separate repos while the session is recorded.

  3. 03

    Break it

    Every build gets hands-on testing: real user journeys, edge cases, deployment, security and code. A polished screenshot doesn't count as proof.

  4. 04

    Prove it

    A transparent rubric, side-by-side results and the raw artifacts (specs, repos, build reports) so you can judge for yourself.

03 The contenders

Three agents. One spec. No head start.

About the agents
  • Claude CodeAnthropic

    Anthropic's agentic coding tool. It works in the terminal, reads the codebase, runs commands and makes changes on its own.

    In EXP-001 Observed: Observed by The Build Hypothesis during the recorded run.

    Claude Code ran through without asking for further input during the run.

  • OpenAI CodexOpenAI

    OpenAI's coding agent. It plans, edits and runs code from the command line, inside a sandbox with configurable approval policies.

    In EXP-001 Observed: Observed by The Build Hypothesis during the recorded run.

    Codex asked for approval whenever it needed to run a new command. Commands it had already been allowed could be allowed again.

  • Cursor CLICursor

    The command-line agent from the makers of the Cursor editor. It runs the Cursor agent headless in any terminal.

    In EXP-001 Observed: Observed by The Build Hypothesis during the recorded run.

    Cursor CLI asked for approval once and offered a "run everything" option, which was selected.

04 House rule

An agent's claim is not a result.

Coding agents write confident reports. Every fact on this site is labeled with where it came from, so "the agent says it works" never gets mistaken for "we tested it and it works".

Read the methodology
  • Spec: A requirement from the versioned specification.
    SpecWhat the challenge required, from the versioned specification.
  • Agent claim: Reported by the agent in its own build report. Not independently verified.
    Agent claimWhat the agent said it did, in its own build report.
  • Observed: Observed by The Build Hypothesis during the recorded run.
    ObservedWhat we saw happen during the recorded session.
  • Verified: Confirmed by independent testing.
    VerifiedWhat independent, hands-on testing confirmed.

05 Up next

Harder challenges are already specced.

See the roadmap
EXP-002
PlannedCreative tool

AI Video Editor

A real browser video editor with a multitrack timeline, AI captions, plain-English edits and a genuine MP4 export.

  • CCClaude Code
  • CXOpenAI Codex
  • CUCursor CLI
Time budget
8h / agent
Spec version
Spec v1.1
Runs
Runs not started
EXP-003
PlannedGame

3D Basketball Game

Five-on-five basketball in the browser, with real ball physics, keyboard control and an AI opponent.

  • CCClaude Code
  • CXOpenAI Codex
  • CUCursor CLI
Time budget
8h / agent
Spec version
Spec v1.0
Runs
Runs not started
EXP-004
PlannedGame

Project Bulletstorm

A stationary 3D lane shooter where you steer an auto-firing squad, recruit soldiers and survive ten levels.

  • CCClaude Code
  • CXOpenAI Codex
  • CUCursor CLI
Time budget
8h / agent
Spec version
Spec v1.0
Runs
Runs not started

06 From the channel

Latest episodes

All videos
Checking the channel for new episodes…

Watch the agents build, break and prove it.

Every experiment becomes an episode: the spec, the runs, the failures and the hands-on evaluation. Subscribe to see the results first.