Why agent-first?

The human moves from CAD operator to orchestrator. Verification becomes mechanical.

Impulse ushers in agent-first engineering — a fundamental inversion of the product development paradigm. AI stops being a command-by-command modeling helper — the copilot era — and becomes an autonomous agent that manages entire multi-step builds: scoping the product, planning the assembly, authoring every part, running physics checks, and committing verified geometry, with minimal human supervision.

01

The shift from operator to orchestrator

Intent-driven architecture

Engineers no longer spend their time driving feature trees, sketching constraints, or debugging mates. They act as engineering directors over the agent — defining problem constraints, functional requirements, and design intent. The system compiles intent into manufacturable geometry.

The factory model

A product is realized part-by-part in parallel goal loops. Each part lands as a reviewable, verified commit — the human reviews outcomes at the boundaries, not keystrokes in the middle.

Outcome over artifacts

Delivery is a working, manufacturable product — DFM-checked, STEP-exportable — not a pile of files. The agent may author and discard dozens of candidate constructions to get there; the capability is the asset, not the clicks.

Verified commits · live
auger · torque path 2.1 N·m ≥ 1.8 spec
02

Radical changes to the development lifecycle

Shrinking work units

Multi-week design revisions collapse into goal-oriented tasks executed in minutes, with human checkpoints at verified commits.

Spec-driven development

Because agent velocity is high, loose requirements become precise, structured contracts: fits, clearances, load expectations, process limits. An agent without rigid constraints accumulates physical debt— parts that don't fit, discovered at tooling, where the fix costs 100× and the launch window doesn't wait. Impulse's constraints are enforced at commit time, so the debt cannot accumulate.

Verification at agent speed

Human design review cannot keep up with agent-speed authoring. Every change is measured against everything it touches — real solvers, invoked mid-reasoning — and everything downstream re-verifies automatically. Testing is continuous and built into the runtime, never a bottleneck the team has to script.

verification loop · replay
03

Changing infrastructure and runtimes

Demoting the GUI layer

Thirty years of CAD interface exists to accommodate human reading and clicking speeds. In agent-first engineering the agent drives the B-Rep kernel and the solver bench directly through APIs — the reasoning loop talks to geometry, not to menus.

Compute-heavy environments

A single design intent triggers dozens of kernel constructions, solver runs, and re-derivations inside a sandboxed realization loop. Only verified geometry ever surfaces. The economics are metered compute, priced against the work.

defineSketch(base)extrude 44 mmshell 2.0 mm✗ min-wall 1.2 at ribre-lower: relieved shellshell ✓fillet 3 mmmeasure(fit) ✓pattern ×6subtract(cutter)measure(clearance 0.82) ✓loft(chute)boolean union✗ non-manifold seamre-lower: decomposed unionunion ✓assignMaterial(PC-ABS)export STEP ✓defineSketch(base)extrude 44 mmshell 2.0 mm✗ min-wall 1.2 at ribre-lower: relieved shellshell ✓fillet 3 mmmeasure(fit) ✓pattern ×6subtract(cutter)measure(clearance 0.82) ✓loft(chute)boolean union✗ non-manifold seamre-lower: decomposed unionunion ✓assignMaterial(PC-ABS)export STEP ✓
04

The structural evolution

DimensionTraditional product developmentAgent-first engineering
Primary human outputDriving CAD feature trees line-by-lineStructuring intent and reviewing verified builds
Development cycleSerialized stage gates, weeks per revisionAsynchronous goal loops, checkpoints at commits
Testing paradigmHuman design review and prototype roundsContinuous measured verification with refusal
Application interfaceGUIs built for human handsA kernel and solver bench driven by reasoning loops

Agent-first engineering hasn't eliminated human engineering intelligence — it has raised it. System architecture, design judgment, and domain expertise remain human-owned.

What changed is who underwrites correctness: the human used to be the verifier of every move; now verification is mechanical, and the human spends their intelligence where it compounds.

This paradigm is live today.

An agent that designs whole products as real manufacturable CAD — and proves every change against physics before it commits.