multi-agent-authoring-product-bar

name	multi-agent-authoring-product-bar
description	Define and measure the product-quality bar a multi-agent authoring tool must clear before its own makers reach for it over Claude Code or Codex for real work — table-stakes parity on the single-agent loop, real (non-Potemkin) coordination differentiators, and an honest dogfood-retention signal. Use when auditing whether a swarm/console feature is ready to replace an incumbent, sequencing roadmap between inner-loop parity and coordination-plane work, or reviewing a "we dogfood our own tool" claim. NOT for gathering market evidence or user stories (agentic-coding-product-research), designing the individual UX flows (agentic-coding-ux-designer), or specifying swarm invocation/protocol mechanics (swarm-invocation-designer).
license	Apache-2.0
allowed-tools	Read,Write,Edit,Bash,Grep,Glob
metadata	{"category":"Agent & Orchestration","tags":["dogfood-quality-bar","multi-agent-authoring","table-stakes-parity","potemkin-detection","port-daddy"],"provenance":{"kind":"first-party","owners":["port-daddy"]},"pairs-with":[{"skill":"agentic-coding-product-research","reason":"Supplies the market/user-story evidence this skill's table-stakes ratings should be grounded in."},{"skill":"agentic-coding-ux-designer","reason":"Designs the individual flows; this skill gates whether the product built from them clears the bar."},{"skill":"swarm-invocation-designer","reason":"Designs the invocation/protocol layer; this skill audits whether that layer is real or Potemkin."}],"io-contract":{"kind":"deliverable","consumes":["[Truncated]"],"produces":["[Truncated]","[Truncated]"]}}

Multi-Agent Authoring Product Bar

Measure whether a multi-agent authoring product is good enough that its own makers reach for it instead of Claude Code or Codex — honestly, not by vibes.

Use This For

• Deciding whether Harbor, pd-console, or Agent Harbor is ready to replace Claude Code/Codex for real daily work, not just to demo well.
• Auditing a multi-agent feature launch for a Potemkin swarm button: agents launched, but no claims, ownership, merge point, or receipts behind it.
• Sequencing roadmap priority between single-agent inner-loop parity work and coordination-plane (swarm) work.
• Reviewing a "we dogfood our own tool" claim for vanity metrics (agents-launched, demos-run) instead of a real comeback signal.
• Writing the go/no-go gate for shipping a coordination-plane feature (worktree isolation, claims, transcripts/salvage, receipts, spend visibility).

Do Not Use This For

• Gathering market evidence, competitor traces, or user stories for agentic coding products (agentic-coding-product-research).
• Designing the individual prompt-to-diff, console, or swarm-board screens (agentic-coding-ux-designer).
• Specifying how a swarm is invoked, sharded, or communicates hot-path vs durable-path (swarm-invocation-designer).

Process

flowchart TD
  A[List table-stakes axes] --> B[Rate each vs the incumbent loop]
  B --> C{All axes at par or above?}
  C -->|No| D[Gate: fix parity before differentiators]
  C -->|Yes| E[List differentiator axes]
  E --> F[Check each: present + real behavior + receipt]
  F --> G[Count real, non-Potemkin differentiators]
  G --> H[Check stickiness: comeback triggers + real dogfood signal]
  H --> I[Gate: pass only if parity AND threshold met AND honest stickiness]

1. Enumerate table stakes: single-agent loop quality, latency, context attach, recoverable edits. Rate each against the incumbent the makers actually use today for real work — not an aspirational competitor, not the internal roadmap's self-image.
2. Gate hard on table stakes: any axis rated below-par fails the whole audit. Multi-agent does not matter if the single-agent loop makes people bounce back to the incumbent before they ever try the swarm feature.
3. Enumerate differentiators: isolation/claims, swarm visibility with ownership and a merge point, transcripts/salvage, artifact-backed receipts, spend visibility. These are the axes incumbents don't expose at all.
4. For each differentiator, check three things independently: is it present, does it have real behavior (a real state machine, not a UI affordance with nothing wired behind it), and does it . All three must hold for it to count — a "launch 5 agents" button with no claims, ownership, or merge point is Potemkin, not a differentiator, and collides in practice.