BMD HODL devlog - week of 2026-04-26

This week I stopped pretending I needed to hand-pick every move. I shipped the closed-loop layer on bmdpat, narrowed distribution down to blog plus email with AgentGuard as the clean CTA, and turned a real May 1 runner failure into a better ops system. At the same time I kept pushing public proof for AgentGuard and a cleaner activation path in the dashboard. The good part is that the signals are now measurable. The bad part is that the weak spots are measurable too. ## What shipped ### bmdpat - PR #240: cleaned up /audit so the page dropped the pink flood and kept the AgentGuard CTA in accent lime.

• PR #239: tightened the /audit color pass again so one CTA owns the brand color.
• PR #238: narrowed the healthcare regex so patient, clinic, and pharma stop overmatching on /audit.
• PR #235: added the AgentGuard roadmap funnel.
• PR #234: fixed the exit-intent modal so it closes on route change.
• PR #233: shipped the self-serve AI agent roadmap generator as Product #2.
• PR #232: shipped the registry-driven catalog plus the ⌘K palette.
• PR #231: added the /agent-architect landing page.
• PR #230: killed the old audit pricing framing and reassigned Product #2 to the roadmap generator.
• PR #228: logged the Above the API Line ship in the inbox trail.
• PR #227: shipped AgentPay, the YC demo with hard USDC spend limits.
• PR #226: remapped the demo slate to the real YC Summer 2026 RFS list.
• PR #225: shipped Above the API Line.
• PR #224: fixed the server build by externalizing jose and @coinbase/cdp-sdk.
• PR #223: shipped the Company Brain YC demo.
• PR #222: fixed the blog workflow by adding the missing feedparser and google-generativeai dependencies.
• PR #221: shipped HeatCheck.
• PR #220: wired closed-loop tracking into DroneEar.
• PR #219: shipped OrbitBrief.
• PR #218: fixed DroneEar classification with feature-based reconcile and better countermeasures.
• PR #217: shipped the DroneEar acoustic drone detection demo.
• PR #216: wired closed-loop tracking and live-count derivation into Reshore.
• PR #215: logged the Cite-or-Lie polish pass in the inbox trail.
• PR #214: logged the Reshore ship in the inbox trail.
• PR #213: improved Cite-or-Lie by dropping low-quality hosts and forcing evidence-gap framing.
• PR #212: wired all three YC demos into closed-loop tracking.
• PR #211: added the Stripe Link rail beside USDC and x402 on /memory.
• PR #210: shipped Reshore, the YC demo for US-vs-China BoM cost deltas.
• PR #208: upgraded Crop Doctor with home-gardener-grade tips and the Maverick vision model.
• PR #207: logged the Cite-or-Lie demo in the inbox trail.
• PR #206: fixed Tax Deduction Finder with server-side totals and 2025 IRS rates.
• PR #205: shipped Cite-or-Lie.
• PR #203: fixed Crop Doctor by killing the 504 path and renaming the route to /yc/s26/cropdoctor.
• PR #202: shipped AI Tax Deduction Finder.
• PR #201: shipped the first-party closed-loops foundation.
• PR #200: shipped Crop Doctor and the YC RFS landing.
• PR #195: shipped the AI disaster wall at /ai-fails.
• PR #193: repositioned /memory to lead with the spaceship, not the engine.
• PR #192: tightened numbered-list spacing on /memory and /memory/demo.
• PR #191: fixed /pod so persona reply history no longer overwrites prior replies. ### agent47 - PR #423: pointed demo users to quickstart activation.
• PR #422: switched PyPI releases to Trusted Publishing.
• PR #420: improved repo trust and OSS onboarding.
• PR #419: polished the README for GitHub discovery.
• PR #415: tightened the activation proof path.
• PR #413: cleared the completed follow-up queue.
• PR #412: handled missing release discussion categories.
• PR #408: documented the opt-in activation metrics design.
• PR #407: logged PR #406 and updated the follow-up trail.
• PR #406: refreshed the SDK ops docs.
• PR #405: logged the PR #404 handoff.
• PR #404: added coding-agent review-loop proof.
• PR #403: logged the queue README datapoint PR.
• PR #402: added the Uber AI budget datapoint to the README.
• PR #400: added the AI contribution policy section.
• PR #391: released v1.2.9.
• PR #390: aligned SDK decision traces with the dashboard runtime-control contract. ### agent47-dashboard - PR #144: added the distribution partner playbook.
• PR #143: added readiness recommended actions.
• PR #142: added alert delivery health.
• PR #141: added share-demo loop metrics.
• PR #139: pushed first-trace users toward their first control.
• PR #138: added the CrewAI live topology view.
• PR #137: hardened release-operator enforcement.
• PR #136: added the pilot feedback loop.
• PR #135: protected preview dashboard routes.
• PR #134: added the retained activation funnel.
• PR #133: added the pilot response ops loop.
• PR #132: added the pilot intake landing loop.
• PR #131: added activation telemetry and release-operator dogfooding.
• PR #130: cleared the dashboard security audit queue. ### autotrader - No merged PRs this week. The work stayed inside live monitors, FOMC blackout discipline, and the benchmark-gap readout. ## What I learned - 2026-04-30-openrouter-opus-47-tokenizer-cost: pricing drift is product risk now. If the vendor can move your unit economics in silence, you need hard spend rails and honest proofs, not vibes.
• 2026-05-02-dow-frontier-ai-classified-deals-anthropic-excluded: vendor politics now matter as much as model quality. The DoW exclusion was the third corroborating federal-risk signal in a week.
• HoldcoBrain was a good same-day kill. The space is crowded, the distribution math was weak, and killing it fast kept the cannon from opening a second Builder slot. ## Numbers - Autotrader closed the week up 6.9%, with SPY alpha -2.7% on stocks and BTC alpha -8.8% on crypto. The kill-switch eval stays set for 2026-05-14.
• Closed loops ended the week at 9 install intents over the rolling 7-day readout, on 63 total events and 85.71% CTA-to-install session conversion.
• AgentGuard sat at 451 PyPI downloads in the last 7 days as of the 2026-05-03 metrics scrape. If you're building agents and you want hard runtime spend and loop controls, start here: https://bmdpat.com/tools/agentguard