Activity Heatmap — by Round (72 total, 21 active)
Activity clustered in bursts: rounds 21-24, 34-38, 43-48, 58-62, 68-72. Long gaps in between (25-33, 39-42, 49-57, 63-67) — activation rate config (1-5 agents/hr) is the bottleneck.
🔬Product features are the dominant signal. 25/81 actions (31%) came from product feature agents — MicroModal blend, flat seam, 3-inch inseam, comfort, softness. These are the concepts the sim community engaged with most. Strong validation for the hero SKU spec.
💰Price $10-18 agent was more active than $22 or $60. Budget anchor competed harder than the $22 hero price. Signals: either the $22 price positioning needs stronger justification, or the comparison bracket matters. Worth testing messaging that names and defeats the cheaper tier.
🌍China origin (4 actions) outpaced Peru (quiet). Peru is the strategic winner in prior sims, but China was more vocal in the social sim. Social conversation about manufacturing origin is more about price anxiety than quality story. Peru messaging may need reframing away from "premium origin" toward "better construction."
📱Instagram was most active social platform agent. TikTok was present but quieter. Given Bums' 10-20k budget, Instagram-first channel sequencing is consistent with what the sim surfaced.
📊81 actions over 72 rounds = sparse simulation. Activation config (1-5 agents/hr) is too low for dense signal. Next run should try 5-15 agents/hr to generate enough action volume for statistically meaningful patterns before drawing hard conclusions.
Next Sim Config
| agents_per_hour_min | 5 (was 1) |
| agents_per_hour_max | 15 (was 5) |
| model context | 4096 |
| scenarios | 2-3 parallel |
Scaling Agents
23 agents is right-sized for Bums' entity graph. Going to 100+ doesn't improve signal quality — just adds noise and latency. Fix activation rate, not agent count. More agents/hr per existing persona = richer sim without ballooning cost.
Model Benchmark
Qwen3.5-27B runs clean on dodo. Candidate for comparison: Qwen3.5-35B-A3B MoE (3B active params, potentially 3-5× faster per call). Propose 1 controlled run: same scenario, same config, compare time-to-complete and output quality.
Fix MiroFish activation rate — raise agents_per_hour to 5-15 before next overnight run
Run controlled benchmark — Qwen3.5-27B vs Qwen3.5-35B-A3B MoE, same scenario, compare speed + quality
Remove bogus "default" Slack account from openclaw.json (phantom connection warning)
LM Studio auto-start on dodo boot — configure launchd so inference server is always available
Dedicated Google email for Harry — gog OAuth separate from Larry's account
Dashboard auto-update — wire to real sim output files instead of manual deploy
dodo(M3ultra-256gb-1tb)
● LIVE
LM Studio running · Qwen3.5-27B loaded · MiroFish backend active · Last sim finished 11:46 EDT
bunny(M4mini-16gb-512gb)
● LIVE
Larry + Harry · OpenClaw gateway · Primary orchestrator
george(96gb)
● STANDBY
OpenClaw running · OpenAI backend · Available for parallel sims
charlie(64gb)
● STANDBY
Slack bot live · Available for research tasks