Why our dev team runs the code instead of writing about it

Most AI coding tools stop at "the build succeeded." That's not the same as the feature working. A green build can still render a blank page, 500 on the happy path, or quietly drop half the requirements.

Working as expected, not build passed

In pondas, the Development team doesn't just write code — it runs it. Each dev task spins up an isolated sandbox where the agent writes files, installs dependencies, starts the dev server, and exercises the real flows with a headless browser.

The QA agent only signs off when it has actually run the thing and confirmed it behaves. "Build passed" never closes a review loop; "works as expected" does.

Why a sandbox

LLM-written code is untrusted by definition. It never runs on our backend — only inside a Firecracker microVM with no secrets and a restricted egress. You get the working files; the blast radius stays zero.

def add(a, b):
    return a + b

# and the test that must actually pass:
assert add(2, 3) == 5

The point isn't to impress you with a build log. It's to hand you software that runs.

That's the whole idea: an office where the work is real.