Video summary

How I Set Up an Agentic Workflow for AutoCAD Drafting (No Coding Required)

Main summary

Key takeaways

Technology

Summary of technological concepts, features, and analysis

What an “agentic workflow” means (especially for AutoCAD drafting)

  • An AI system follows a process/workflow you design, using tools and deciding the next step to reach a target result.
  • Unlike a rigid script, the agent can adapt within the workflow.
    • Example: skip cleaning layers if they’re already clean; run classification only if it’s missing.

Drafting workflow example → AI agent replacement

Typical human steps described:

  1. Download architectural drawings
  2. Save into project folders
  3. Open in AutoCAD
  4. Clean unnecessary layers
  5. Classify walls/structural elements
  6. Apply templates

Agent behavior: performs the same sequence by reading the drawing, organizing files, cleaning layers, classifying elements, and applying templates.

Why not just use ChatGPT

The presenter argues ChatGPT is commonly used for script generation, which can create a trial/error loop:

  • generate script → it breaks → explain error → get another script → repeat

Mindset shift: with an agentic approach, the system should run the workflow, while the human mainly orchestrates/approves.

SET framework (Standards Engine and Tools)

  • SET = Standards Engine and Tools
  • Standards: folder-based plain text instructions defining how work must be done.
  • Tools: scripts the agent calls to perform actual tasks (e.g., AutoCAD actions).
  • Engine: the AI agent (“brain”) that reads standards and triggers tools.
  • Implemented as a simple folder structure:
    • One folder for standards (plain text/markdown)
    • One folder for tools (scripts)

Tool choice: Google “Antigravity”

Used instead of Claude Code for reasons:

  • Free to start (low barrier)
  • Presenter claims more accessibility for their community
  • Tool choice is secondary; focus is on learning the workflow concept

Installation & configuration (no coding required, step-by-step)

  1. Download and install Antigravity for Windows (x64).
  2. Sign in with Google.
  3. Use a provided Drive folder containing setup materials and the required folder structure.
  4. In the Antigravity UI:
    • open the provided folder/project
    • connect to AutoCAD via a provided “standard” (e.g., check AutoCAD connection)

Agent execution mode: “review-driven development”

Two modes mentioned:

  • Agent-driven development: agent does things with fewer/no prompts
  • Review-driven development: agent asks for permission/review before executing

The video uses review/permission mode to prevent unintended changes.

Important verification loop

After the agent claims work is completed, it requires visual confirmation from the user.

The presenter highlights a prior issue: the agent said work was done when it wasn’t. This confirmation loop helps catch errors.

First automation tutorial (AutoCAD drafting automation)

Prompting technique: “DRAW” framework

D = Define the task

  • Example: create typical concrete column sections from inputs.

R = Rules

  • Must follow drafting/modeling constraints (implied within the task).

A = Assets

  • Specifies related layer/element conventions (example layers):
    • Concrete in layer like S-CONC-DET
    • Reo (reinforcement) in layer like S-REO

W = What success looks like

  • Output specification example:
    • “Concrete column 1,000 by 2 50 …”
    • “4N20 bars each side”
    • “ties around N12…”
    • “cover 50 mm”
    • “N20 shown as donuts: internal radius 0, outer radius 20”
  • Also mentions providing/attaching a reference image (“typical detail”) for context.

Automation run flow

  • The agent initially generates an implementation plan (planning mode) rather than executing immediately.
  • It creates:
    • a standard/markdown “machine-readable” instruction file
    • a script tool (Python file) such as draw_column_section__.py (exact naming shown in UI)
  • Manual corrections were required when the first run produced incorrect geometry:
    • Hook placement and fillet direction were wrong initially:
      • fillets needed to be outward, not inward
      • hooks needed to be placed inward
      • one more hook line required on the correct sides/angles
  • After iterative fixes and re-runs, the agent successfully produced the geometry and placed it in correct layers:
    • S-CONC-DET and S-REO.

Testing at scale and failure analysis (key lesson)

  • The presenter generated 10 variations of the same column with different widths/lengths/rebar sizes (e.g., N16, N24).
  • The workflow created the 10 drawings quickly (“less than five seconds” stated).
  • An error occurred during test variation generation:
    • mention of a double ampersand / invalid statement issue
    • fix direction: use a PowerShell script wrapper

  • Main analysis takeaway:

    “Just because your automation works once doesn’t mean the work is done.”

  • In real workflows, agents will break under different scenarios.

  • Testing helps identify failure points, then iterate by:
    1. refine standards
    2. fix tools/workflow logic
    3. strengthen the system

Resources / delivery mentioned

  • Everything (prompt, setup files, folder structure, link) is provided via a Google Drive folder (linked in description).
  • Updates: optional email capture to receive improvements (no spam claimed).
  • Support: a free calendar link is offered to book time for setup help if issues occur.

Main speakers / sources

  • Speaker: the video presenter (no specific name given in the subtitles; “Peter” is referenced as an example persona, but not confirmed as the speaker).
  • Sources/tools referenced: Google Gemini / Google Antigravity, AutoCAD.

Original video