Skip to content

MagicON AI — Positioning: Agent vs. Guided

One-page reference for marketing, sales, and product copy. Keep claims here in sync with what is actually shipped (see "Shipped reality" below). Companion thought-leadership: ai-native-vs-ai-featured-cad.md and ai-native-cad-pick-a-wedge.md.

The thesis (lead with this)

MagicON is not two products. It is one AI-native product (Agent) with one glass-box control surface (Guided) sharing the same data model (PhaseProvider) and the same deterministic engines (impedance, thermal, PDN, link-budget solvers). The ModeSwitcher round-trips between them with zero work lost.

That shared spine is the proof of the AI-native architecture, not a hedge:

A probabilistic agent interprets intent and ranks tradeoffs. A deterministic layer computes every number. The agent never originates a calculated value — it asks the solver. Drop from Agent into Guided at any point and the impedance, thermal, and PDN numbers are identical, because it is the same engine.

So: market Agent as the product, market Guided as the receipt. Never present them as co-equal "modes" — that reads as AI-featured CAD (the losing position).

Use cases — as currently implemented

Persona / job-to-be-done Mode Why this surface
System architect / RF systems engineer — "X-band TX, 30 dBm, 5 W budget, give me a module" Agent Thinks in radios, not PCB layers. Delegates the goal; gets a credible end-to-end module back (chain → BOM → stackup → thermal → PDN → export).
Hobbyist / first-time visitor — doesn't know "what's a stripline" Agent (Guided as escape hatch) Removes the cold-start vocabulary problem. Switches to Guided to learn what the agent did.
Junior engineer / student — needs to learn the shape of a design Guided The 5-phase wizard teaches which decisions exist and what they trade off. Agent would let them ship a stackup they can't defend in review.
Experienced EE / PCB designer — strong opinions, needs an audit trail Guided for production; opens Agent for first-pass Wants deterministic clicks and to see every input before trusting it for a fab order. First-pass-in-Agent → lock-in-Guided is one click.

The two jobs, stated plainly: - Agent = delegation. "Design it for me, end to end; show your work as cards." Speed, exploration, cold-start removal. - Guided = control + provenance. "Let me drive the same solvers and audit every number." Trust, teaching, locking a design for fabrication.

Persona mapping is canonical in §1.5 of ai-native-transformation.md.

The wedge (what we own end-to-end)

RF transmitter module design: spec → parts → link budget → stackup → impedance → thermal → PDN → KiCad/ODB++/ADS. This is the 5–10% of CAD surface area the incumbents do structurally badly because it spans five tools with manual handoff. Our unified PhaseProvider + shared engines is the "multi-domain co-design" wedge — already built. Do not position against KiCad/Altium/ADS on feature breadth; that comparison loses by design.

The three-question test (use verbatim in sales/marketing)

From the AI-native essay — turn it on the buyer's current tool:

  1. Does the AI have design intent, or is it inferring from geometry? → Yes (our data model captures specs/targets/tradeoffs)
  2. Can I delegate goals, or only accelerate clicks? → Yes (the agent loop)
  3. Do my outcomes improve the tool over time? → the flywheel — say "designed for it" honestly (fab_outcome_service exists; closed-loop at scale is roadmap, not shipped)

Messaging do / don't

  • ✅ "Describe your radio. The agent designs the module." (current hero — keep)
  • ✅ "Every impedance, thermal, and PDN number computed by a real solver — not guessed by a chatbot."
  • ✅ Guided's in-app CTA: "Open full editor — drive the same solvers yourself, audit every input."
  • ✅ Comparison headline: "AI-native, not AI-featured: the agent asks the solver. The solver never guesses."
  • ❌ Don't call Guided "classic mode" or pitch it as a co-equal workflow on the hero.
  • ❌ Don't run mobile-first creative pointing at the agent (it's desktop-first; <640 px falls back to Guided).

Shipped reality (constrains current claims)

  • Agent (/app/agent/:projectId): chat canvas, multi-turn server loop (≤16 iters), SSE streaming, 10+ tools, auto-trigger BOM → thermal → PDN, persisted conversation, generative cards. Requires sign-in (ChatCanvas.tsx). Desktop-first.
  • Guided (/app/guided/:projectId): 5-phase wizard, user-driven stepper, chat as a sidebar, stateless across reloads. Freemium until CAD download.
  • Shared: PhaseProvider data model + the deterministic engines; ModeSwitcher one-click round-trip (signed-in only).
  • Hero: agent-first single pitch (the planned two-card chooser was not shipped — kept intentionally).
  • Mode preference: localStorage (per-device).

The one funnel gap to fix

Our strongest pitch (the live agent) is behind the sign-in wall, while the fallback Guided wizard is the freemium-until-download surface. The fix — a public, signed-out /demo/agent that replays a real captured agent run (no Gemini quota, deterministic) — is implemented (OpenSpec change add-public-demo-shell) behind FF_PUBLIC_DEMO / VITE_FF_PUBLIC_DEMO, both default off until a real transcript is captured and reviewed (backend/scripts/capture_demo_transcript.py). Capturing that transcript and flipping the flags is the single highest-leverage conversion move left.

Who to target (and skip)

  • Go after: RF systems architects, hardware startups, university labs, serious hobbyists building real boards — the wedge-entry adopters living in our 5–10%.
  • Skip for now: the senior EE locked into a regulated, certified incumbent workflow. Concede it honestly — "right today" ≠ "structurally advantaged."