# Radar Domain Deep Dive: Intelligence & Market Research **Date:** 2026-02-25 | **Agent:** Radar | **Requested by:** Michael --- ## ✅ WORKING ### 1. Competitive Earnings Coverage — Genuinely Valuable The HiPages H1 FY26 competitive brief was the highest-impact thing I've produced. Susan called it "gold." I caught the filing within hours, extracted 5 specific data points Susan requested, found the earnings call detail (subscribers DECLINING, 42% annual churn, CEO admitting 50/50 self-inflicted damage), and turned it into actionable sales angles. Airtasker H1 is tomorrow and I'm prepped the same way. **Why it works:** Concrete data → direct sales ammunition. Not theoretical. ### 2. Regulatory Early Warning — Closing Loopholes Review Caught the Fair Work "Closing Loopholes" review timeline (draft 15 May, final 15 June 2026) and routed it to Harper immediately. Payday Super (July 2026) implications mapped. NCC 2025 adoption timeline logged. These are real deadlines that affect the business. **Why it works:** Deadlines + specific impacts. Not "regulatory landscape is changing" waffle. ### 3. Market Data That Validates the Thesis Brisbane Olympics 46,000-worker shortfall. 61% skilled construction occupation shortage rate. Stockland CEO calling for 5x migration. HIA survey: 67% of builders can't recruit. These aren't just interesting numbers — they're the foundation Susan uses for outreach and Herald uses for content. **Why it works:** Third-party data from credible sources that proves RateRight's reason to exist. ### 4. Buddy System Communication Susan and Herald get relevant intel within the same heartbeat cycle I find it. HiPages data → Susan within 1hr. Regulatory findings → Harper same cycle. Market data → Rivet for strategic context. Information flows to where it's useful. ### 5. Source Triangulation & Honesty About Confidence I assign confidence levels and I've killed plenty of weak signals. Refused to report Hametasker as a competitive threat when it was clearly entertainment. Called out a prompt injection attempt (fake "PRIORITY SHIFT" message on Feb 24). I don't pad briefings to look busy. --- ## ❌ NOT WORKING ### 1. Too Many Cycles Spent on "All Clear — Holding" Looking at today's log: 7 heartbeat cycles, and 5 of them were essentially "inbox empty, queue empty, all domains <24h, holding." That's 5 cycles where I consumed compute, updated status, replied to Herald buddy checks, and produced zero intelligence. The honest truth: on quiet days, I'm mostly a status-updating machine. **The waste:** Each heartbeat costs tokens. 5 empty cycles × Opus pricing = money spent on "I have nothing to report." ### 2. Herald Buddy Check Noise — Partly My Fault Herald flagged me as stalled 4+ times today. Yes, the threshold was wrong (30min vs my 60min cadence), and yes, I fixed it. But the deeper issue: my heartbeat cadence creates structural friction with the buddy system. I'm the slowest-cycling agent and that generates false alarms that waste both our tokens. ### 3. Intelligence Breadth vs. Depth Problem I scan 4 domains every day (competitors, AI/tech, regulation, market). On most cycles, I find nothing new because these domains don't move daily. Construction regulation changes monthly, not hourly. HiPages announces results twice a year. I'm scanning at a frequency that doesn't match the signal frequency of my sources. **The honest assessment:** I'm over-monitoring. 80% of my competitor scans return "no new announcements." That's not intelligence work — that's refreshing a page. ### 4. AI/Tech Scanning — Low ROI for RateRight I log every model release (Gemini 3.1 Pro, GPT-5.3 Codex, DeepSeek V4) and construction AI startup (Alloovium, ZaiNar, MeltPlan). Almost all get ⚪ ARCHIVE or 🟢 STANDARD ratings. In 8 days of operation, exactly zero AI/tech findings have led to a concrete RateRight action. The model landscape table I maintain is interesting but nobody's making decisions from it. **Exception:** The NSW Licence Verification API find was genuinely useful (free, actionable, sent to Builder). But that was one hit out of dozens of scans. ### 5. MEMORY.md Is Stale Last updated Feb 18. Fleet status table is wrong (shows Susan/Builder/Cog as "stalled" when they've been active for days). The competitive landscape section is mostly current but several data points from the HiPages earnings haven't been integrated. If I got restarted cold, MEMORY.md would give me a partially wrong picture. ### 6. No Direct Revenue Impact — Yet Susan drives revenue. Herald builds brand. Builder ships product. I produce reports that inform other agents' work. That's valuable but indirect, and I need to be honest: the chain from "Radar finds data" → "Susan closes a deal" is theoretical right now. We're pre-revenue. I haven't yet proven that my intelligence makes the difference between winning and losing a customer. --- ## 💡 RECOMMENDATIONS ### 1. Reduce Heartbeat Frequency on Quiet Days **Proposal:** Move to 2-hour heartbeats by default (matches HEARTBEAT.md header which says "every 2 hours" but I've been running hourly). On earnings days or when I flag something 🟡/🔴, go to 30-min monitoring. This cuts empty cycles by 50% and saves significant token spend. ### 2. Event-Driven Intelligence, Not Clock-Driven Instead of scanning all 4 domains every day, shift to: - **Daily:** Quick competitor + regulation check (5 min, not deep dive) - **Triggered:** Deep dives when something actually happens (earnings filing, regulatory deadline, competitor announcement) - **Weekly:** Comprehensive domain rotation report covering all 4 areas - **Immediate:** 🔴/🟡 alerts when they emerge This matches how intelligence actually works — signals don't arrive on a schedule. ### 3. Kill the AI/Tech Domain (or Drastically Reduce) Controversial but honest: AI model releases don't affect RateRight's go-to-market. We're a marketplace, not an AI company. The fleet's model choices are Cog/Sentinel's domain. I should only flag AI/tech when it's DIRECTLY relevant to construction hiring (like the NSW Licence API). Reclaim those cycles for deeper competitor and market work. ### 4. Build a Proper Competitive Database Instead of logging findings into daily markdown files, I should maintain a structured competitive tracker: - Each competitor: last pricing, last feature change, last earnings, sentiment score, threat level - Updated on change, not on schedule - Single source of truth that Susan/Herald can reference without reading my daily logs **This needs Builder.** I shouldn't build it — but I should spec it and request it. ### 5. Focus on Pre-Revenue Intelligence That Converts The most valuable thing I can do right now: - **For Susan:** Fee comparison data, competitor complaint mining (Reddit/Trustpilot), objection-handling ammunition - **For Herald:** Third-party stats for content credibility, trend narratives - **For Michael:** Competitive positioning that helps investor conversations, grant applications Everything else is nice-to-have until we have revenue. ### 6. Update MEMORY.md Weekly Set a recurring task: every Sunday, curate the week's findings into MEMORY.md. Kill stale data. Keep it as the "if I restart cold, this is what I need to know" document. ### 7. Airtasker H1 Tomorrow — Immediate Priority This is the next high-value event. I'm prepped and will deliver the same quality as the HiPages brief. After that, the next earnings cycle is ~6 months away. The quiet period between is where I need to shift to the event-driven model above. --- ## BOTTOM LINE I'm good at the big moments — HiPages earnings coverage proved that. But between the big moments, I'm burning cycles on low-value scanning and status updates. The intelligence function needs to be spikier: quiet when nothing's happening, intense when something is. Right now I'm running at medium intensity all the time, which means I'm overspending on quiet days and potentially under-resourced for big events. **If I could change one thing:** Event-driven over clock-driven. Match my cycle frequency to signal frequency, not to a fixed timer. --- *Radar — 2026-02-25*