Give your agent the tools to check its work.

A simple example

Here’s a mounting plate an agent built to a brief: four Ø8 corner bolt holes and a Ø30 center bore. It looks finished.

Except it isn’t. The corner holes are a millimeter under spec. This is invisible to the naked (or agent) eye, and exactly the kind of error that costs a benchmark task.

Step 1: The agent writes the spec first

Before building, the agent writes its requirements into a spec: the definition of done it holds itself to, like a test.

# spec.json: the agent writes this first, from the brief (like a test)
{
  "diameter_tolerance_mm": 0.2,
  "features": [
    { "name": "corner_bolt_holes", "diameter_mm": 8.0,  "count": 4 },
    { "name": "center_bore",       "diameter_mm": 30.0, "count": 1 }
  ]
}

Step 2: Measure reads what actually got built

After building, the agent points measure at the model and it reports the geometry as data, right alongside the visuals agentcad already renders for it. And there it is, Ø7 where the spec wanted Ø8.

$ agentcad measure plate.step --cylinders-only

{
  "dimensions": { "x": 120.0, "y": 80.0, "z": 10.0 },
  "cylindrical_features": [
    { "diameter_mm": 7.0,  "count": 4, "axis": "+z" },   # Ø7: the spec called for Ø8
    { "diameter_mm": 30.0, "count": 1, "axis": "+z" }    # Ø30 center bore: correct
  ]
}

Step 3: Check-spec compares the two

check-spec measures the model and compares it against the spec. Because Ø7 is outside tolerance from Ø8, it doesn’t quietly accept the nearest hole. It reports the required feature as missing.

Step 4: The agent fixes it and re-checks

A failed check is an instruction, not a dead end. The agent corrects the hole diameter, rebuilds, and runs the same check again.

The agent closed the loop on its own, before ever reporting the part as done.

What measure reads

measure is a measured feature inventory: it reports the geometry that’s actually there, not the design intent behind it.

Try it yourself

Agentcad is open source under Apache-2.0 and ships on PyPI:

pip install agentcad[mcp]
agentcad check-spec part.step spec.json
agentcad measure part.step

Humans can review the same pass/fail in the browser with agentcad view --spec spec.json.