Two bets under one brand. Adam confirmed May 13.
Bet A shipped Thursday for in-market validation. The Validation Framework runs parallel by intent; measurement is sequenced — Bet A reads from launch, Bet B's in-product A/B fires once Bet A's audience is acquired. Triangulation gate w/c 1 Jun evaluates Bet A; Bet B axis follows weeks 4–5.
Agents hub
Infrastructure + agent marketplace. Run any open agent (OpenClaw, Anton; Hermes + NanoClaw via SSH/CLI w/c 18 May) on MindsHub.
Cowork UX
The bet: Anthropic Cowork has the best UX in the category; MindsHub extends it to agents that lack one. Offered per-agent inside the hub.
"Anton is no longer a first-class citizen of Cowork — it's one backend among equals." — Adam, All Hands May 12
Outcome matrix
Triangulation gate w/c 1 Jun (staged). Bet A axis evaluated first; Bet B axis evaluated in weeks 4–5 as the in-product A/B lands. David runs synthesis; Jorge + Adam sign off.
Six layers, parallel by intent. Measurement sequenced.
Layer 1a tests Bet A (hub adoption); Layer 1b tests Bet B. Measurement fires from w/c 18 May; Bet B's in-product A/B holds until Bet A's signal lands ~end of week 3. Layers 0, 2, 3a, 3b feed both bets. Click a layer for detail.
Feeds: Layers 1a, 1b, 3a.
Owner: David. When: w/c 18 May.
Ticket: ENG-53
Measures: hub-feature engagement (Model Router usage, multi-agent runs, model switching), D1/D7 retention, "I came for X" answer pattern.
Pivot signal: users sign up for the Cowork experience → trigger Bet B.
Owner: David. When: w/c 18 May (recruitment starts post-launch).
Tickets: ENG-55 (interviews) + ENG-50 (PostHog instrumentation)
Supporting test: comparative UX eval. Cowork-wrapped vs native clients (OpenClaw, Hermes, NanoClaw) vs Anthropic Cowork (upper bound).
Measures: A/B selection rate, retention by variant, completion time, qualitative preference.
Owner: David (paired with Adam on A/B design once Bet A signal lands).
When: UX eval w/c 18 May; in-product A/B w/c 1 Jun.
Ticket: ENG-56
Pass criteria: within X% of frontier quality at Y% lower cost. Thresholds defined before running.
Outcome rule: failure drops the feature; doesn't kill Bet A (hub stands without auto-routing).
Owner: David + Max A.
Ticket: ENG-52
Launch reality: $9.95 + 5M tokens ≈ 12 Anton queries. Recycled HubSpot top-up form bridges launch; auto top-up wires in with Metronome. Top-up vs cancel rate is a leading WTP signal.
Margin floor: open — Stripe + Metronome cost-data feeds the decision.
Applies to: both bets.
Owner: David + Adam.
Ticket: ENG-57
Depends on: Metronome + billing instrumentation live.
Applies to: both bets.
Owner: David. When: starts w/c 25 May (once billing live); first read w/c 1 Jun.
Ticket: ENG-58
Four weeks from launch to decision.
Each sprint card shows the headline outcome plus the items starting that week. Live sprint state and per-ticket status live in the Scope Document and the Weekly Learning Ledger pages, not here.
- Launched Thursday on mindshub.ai
- OpenClaw + Anton wired to MindsHub model router
- PostHog identity binding from launch
- Personal AWS migration + domain cutover (P0)
- Top-up form available on quickStart + billing/usage
- A1 measurement scaffolding (cohort + session + return-rate)
- Production code freeze Mon noon → Wed noon for Tue board meeting; credibility-gap work merges to staging in the interim.
- PRs mindsdb/minds#277 + #278 merging this week (conflicts resolved Mon morning by Lucas; staging target — promotes to prod after code-freeze lift)
- Layer 0 sentiment scan
- Layer 1a hub-value interviews start
- Layer 1b comparative UX eval (Cowork-wrapped vs native vs Anthropic Cowork)
- Layer 1b in-product A/B fires on acquired audience
- Layer 2 Model Router benchmark — fires once #278 merges; otherwise slips to w/c 25 May
- NanoClaw + Hermes accessible via SSH/CLI (no UI; UI mapping deferred — bigger project; Lucas + Zoran)
- Hermes production wrap (SSH proxy hardening; NanoClaw hardening follows w/c 25 May)
- Auto top-up via Metronome billing instrumentation
- Stripe + Metronome cost data → margin-floor call — depends on #277 merging; token recording must be correct first
- Metering surface, Metronome schema tags
- Kill-switch threshold set (X + N defined on paper; David + Adam)
- BYOB-tier WTP modelling (Layer 3a survey + competitor data; Adam)
- Per-campaign attribution wired (UTM + PostHog acquisition_campaign property; David + Lucas) — TBC Lucas capacity
- Per-campaign interview cohort design + outreach copy (paired with Costa's campaigns) — TBC Costa sync
- Layer 3b price-variant experiment starts (Statsig variants on live cohort)
- First read: Hermes cohort behaviour visible — informs Layer 1a hub-value reads + Walk 1 prep
- First read: Layer 3a Metronome-event signal (top-up vs cancel rate, allocation exhaustion timing)
- First read: Margin-floor recommendation (Costa 5% / David 20% resolves against real numbers)
- First read: Per-backend cohort segmentation (OpenClaw / Anton / Hermes / NanoClaw / BYOB comparable in PostHog)
- First read: Kill-switch math runs against real numbers (per-backend CAC-clearing ratio vs threshold X)
- First read: BYOB-tier WTP recommendation against margin floor (Adam)
- First read: Per-campaign cohort segmentation (which Costa campaign delivered which cohort behaviour; feeds Layer 1a + 3a reads) — TBC Costa sync
- NanoClaw production wrap
- Canonical API v0.1 locked
- Per-campaign interview cohort in flight (~20 power users from Costa's funnel, segmented per-campaign)
- Bet A signal accumulation (Layer 1a interviews continuing)
- Triangulation gate — Bet A axis
- Synthesis: Layers 1a + 2 + 3a + 0 → memo
- Jorge + Adam sign off on Bet A outcome
- Bet B axis follow-up (weeks 4–5 as Layer 1b A/B lands)
The triangulation gate.
A staged decision moment, not a single meeting. Bet A axis decided w/c 1 Jun from Layers 1a + 2 + 3a + 0. Bet B axis follows weeks 4–5 as Layer 1b's in-product A/B lands. David synthesises; Jorge + Adam sign off. Output: a triangulation memo against the outcome matrix above + a Bet B follow-up note.
What could invalidate the bets.
Survival-of-bet concerns that persist across releases. Sprint-execution risks (cutover, sequencing, single-backend timing) live in the Scope Document and the Weekly Learning Ledger.
Anthropic ships multi-agent Cowork
If Anthropic ships multi-agent Cowork — the same Bet B MindsHub is launching — that bet is neutralised overnight. Bet A is less exposed: competes with HuggingFace Spaces / Replicate, not Anthropic.
OpenClaw upstream dependency
Steinberger joined OpenAI; stewardship moved to a foundation. A licence shift compromises Bet A — OpenClaw is the first agent live on the hub, and the hub thesis depends on a project we don't control.
Persona unsettled (resolves via Layer 1a)
Primary persona: technical enough to want OpenClaw, not savvy enough for CLI — already shopping for Claude / ChatGPT alternatives. Pivot signal if Cowork's UX, not the hub features, becomes the point of origin — a Bet B win. Layer 1a watches for it.
Experimentation spend eroding margin
Cost-discovery + model-experimentation spend outpaces revenue. The Model Router (Layer 2) is the structural lever; until it validates, every cohort scaled adds margin loss. The kill-switch threshold is the discipline that caps the bleed; its absence is itself a risk.
Pricing model unsettled — seat, slot, or consumption?
Three shapes the live cohort responds to differently. The choice shapes what Layer 3a/3b measures, what the kill-switch math runs against, and BYOB-tier pricing. Margin floor (~5%–20%) sits inside whichever shape we pick.