From plain English to validated equipment code.

The engineer types what they want. No G-code, no Python, no SDK calls — just a sentence.

$ agent.prompt:

"Run a 4-corner probing routine on the reference fixture,
 compute XY offset and tool-length offset, and write the
 corrected origin to G54. Halt and alert me if any probe
 deflection exceeds 0.2 mm."

The agent queries the vector store for documents that match the intent — vendor manuals, internal SOPs, fixture drawings, prior calibration logs. Grounding is explicit; nothing is hallucinated.

$ rag.retrieve("probing + tool-length offset + 3-axis VMC")
  loaded 8 contexts in 2.4 s ·  embed dim 1024 · top-k 8

  ✓ G-Code Programming Manual — Probing Cycles      [vendor]
  ✓ Tool-Length Offset Procedure                    [SOP]
  ✓ Reference Fixture Drawing #RF-204               [CAD]
  ✓ Machine Safety Boundaries — Spindle Region      [safety]
  ✓ Calibration Run History (last 6 mo)             [runs]
  + 3 more

Output is real code, not pseudocode — typed against the retrieved manual, parameterised from the prompt, runnable by the executor in the next stage.

# auto-generated · agent run f4a1 · 14:23:01
from cnc import Controller
from calib import compute_offset, write_g54

mill = Controller(host="10.0.4.12", auth=KEY)

# safe approach
mill.send("G00 G54 X0 Y0 Z25")
mill.send("G31 Z-20 F500")        # probe down (tool-length ref)
z_ref = mill.read_probe()
mill.send("G00 Z25")

# 4-corner XY probing
probes = []
for x, y in [(-50,-50),(50,-50),(50,50),(-50,50)]:
    mill.send(f"G00 X{x*0.9} Y{y*0.9}")
    mill.send(f"G31 X{x} Y{y} F200")
    probes.append(mill.read_probe())
    mill.send("G00 Z25")

dx, dy = compute_offset(probes)
write_g54(mill, dx, dy)
print(f"applied G54: dX={dx:+.3f} dY={dy:+.3f}")

Before a single command leaves the workstation, the script runs against a virtual mill. Each command is checked against the equipment's allowlist and the operating envelope.

$ sandbox.run(script)

→ simulator boot ............................ 0.4 s
→ allowlist check:
   ✓ G00 (rapid)                — allowed
   ✓ G31 (probe)                — allowed
   ✓ G54 origin-write          — allowed
   ✓ all Z values > -25 mm     — within safety envelope
   ✓ spindle-on commands       — none requested  (correct for probing)

→ virtual run:
   ✓ 4 / 4 probe points succeeded
   ✓ no collision predicted
   ✓ computed offset: dX=+0.140 mm  dY=-0.080 mm
   ✓ estimated runtime: 2 m 47 s

PASS — ready for hardware

The operator confirms in person. Once approved, the script flows down through the runtime stack — each layer signs off before the spindle moves.

The agent renders a self-contained report — chart, table, log location — and writes it to the validation team's shared store. No screenshotting, no spreadsheet wrangling.

$ dashboard.generate(run="2026-05-23_calibration")

CALIBRATION REPORT · 2026-05-23 14:27
─────────────────────────────────────
  Fixture          RF-204 (reference)
  Probe points     4 / 4 successful
  Max deflection   0.08 mm  (threshold 0.2 mm · headroom 60%)
  Computed offset  dX = +0.140 mm   dY = -0.080 mm
  Z reference      +0.000 mm  (unchanged)
  Applied to       G54  (active WCS)

  ⬢ visualisation     [4-corner offset plot, deflection bars]
  ⬢ audit log         audit-log/2026-05-23_1424.signed
  ⬢ stored at         results/2026-05-23_calibration.json

DONE  ·  ready for next shift