---
created: 2026-03-12
source: Rivet
tags: [agent-archive, rivet]
---

# Monthly Review: February 2026

*Compiled by Rivet (subagent) — March 2, 2026*
*Data sources: 15 daily memory files, 13 nightly strategic reviews, 9 days of session cost data, fleet state, decisions log*

---

## 1. COST & TOKEN USAGE ANALYSIS

### Measured Period: Feb 20–28 (9 days with session-level cost data)

| Date | Cost (USD) | Sessions | Messages |
|------|-----------|----------|----------|
| Feb 20 | $4.52 | 34 | 193 |
| Feb 21 | $9.36 | 96 | 564 |
| Feb 22 | $8.30 | 99 | 535 |
| Feb 23 | $7.55 | 104 | 538 |
| Feb 24 | $10.33 | 103 | 609 |
| Feb 25 | $9.01 | 104 | 581 |
| Feb 26 | $9.61 | 105 | 576 |
| Feb 27 | $5.73 | 103 | 544 |
| Feb 28 | $7.33 | 107 | 571 |
| **Total** | **$71.74** | **855** | **4,711** |

**Pre-Feb 20 archive (Feb 1–19 sessions available):** $11.10 across 71 sessions, 599 messages.

**Estimated full-month total:** ~$83–90 (within $100/month budget but tight).

### Cost by Model (Feb 20–28)

| Model | Cost | % of Total | Messages | Cost/Msg |
|-------|------|-----------|----------|----------|
| Claude Sonnet 4 | $49.39 | 68.8% | 3,072 | $0.016 |
| Claude Opus 4.6 | $18.58 | 25.9% | 395 | $0.047 |
| Claude Opus 4.5 | $3.73 | 5.2% | 48 | $0.078 |
| GPT-4o-mini | $0.04 | 0.1% | 18 | $0.002 |
| Kimi K2 | $0.00 | 0.0% | 1,118 | $0.000 |
| **Total** | **$71.74** | **100%** | **4,711** | **$0.015** |

### Key Cost Insights

1. **Sonnet is the workhorse (68.8% of cost, 65% of messages).** Used for all cron jobs, heartbeats, and most automated tasks. Good cost/capability balance.
2. **Opus 4.6 used sparingly but effectively (25.9% of cost, 8.4% of messages).** Reserved for Michael's direct conversations and complex decision-making. 3x the per-message cost of Sonnet but justified for strategic work.
3. **Kimi K2 is essentially free labor (1,118 messages at $0).** Used for 11 of 19 cron jobs. However, **Kimi hallucinated fabricated data on Feb 23** (fake grant submission in evening brief), leading to the upgrade decision. Free but dangerous.
4. **Cache read dominance:** 84M tokens cached for Sonnet, 17.6M for Opus. Cache strategy is working — input tokens are negligible (33K Sonnet, 1.3K Opus) vs cache reads. This keeps costs far below what they'd be without caching.
5. **Daily run rate: ~$8/day.** At ~$240/month extrapolated, this significantly exceeds the $100/month budget. The Feb 20–28 period was high-activity (launch prep + crisis response), which inflates this.

### Cost Efficiency by Task Type (Estimated)

| Task Type | Model Used | Approx. Cost/Day | Value Assessment |
|-----------|-----------|-------------------|-----------------|
| Heartbeats/health checks | Sonnet | ~$2–3 | Medium — necessary but could be cheaper |
| Strategic reviews (nightly) | Sonnet | ~$1–2 | HIGH — identified launch-blocking blind spots |
| Michael conversations | Opus 4.6 | ~$2–3 | HIGH — direct business value |
| Fleet coordination | Sonnet/Kimi | ~$1–2 | Medium — essential but process-heavy |
| Cron jobs (briefs, audits) | Kimi/Sonnet | ~$0.50–1 | Mixed — Kimi too unreliable for customer-facing |

---

## 2. BIGGEST WINS

### Win #1: National Launch Strategy Decision (Feb 20)
Michael won a 20-minute strategic debate against Rivet's textbook "Sydney-first" advice using 30 years of construction industry knowledge. The insight — construction word-of-mouth is cross-city, not local — fundamentally changed the go-to-market approach. This is the most consequential strategic decision of the month.

### Win #2: DNS Migration & Professional Domain (Feb 23)
Michael successfully migrated rateright.com.au from GoDaddy to VPS during a 90-minute evening window. Professional domain achieved for launch — eliminates the credibility gap of a subdomain.

### Win #3: Agent Identity Crisis Resolved (Feb 23)
Builder discovered and fixed a critical memory infrastructure bug: 5 agents were loading Builder's identity files instead of their own — they literally thought they were Builder. This was invisible but catastrophically damaging to fleet coordination.

### Win #4: $50 Flat Fee Model Validation (Feb 19–24)
Radar's competitive intelligence confirmed: **NO competitor offers unlimited duration flat fee.** All charge percentage or monthly subscriptions. "Pay once, hire for years" is genuinely differentiated. 85–90% cost advantage vs Yakka Labour's hourly model.

### Win #5: Fleet Excellence System & Challenge Culture (Feb 24)
Michael identified the "yes-man problem" — agents were agreeing instead of thinking critically. Challenge culture directive deployed to all agents. Self-audit protocol established. This is a structural improvement that compounds daily.

### Win #6: Worker Acquisition Landing Page (Feb 28)
Builder delivered `/join` page in under 2 hours — zero-distraction conversion page targeting WHV/backpackers. Directly unblocked Susan's hostel distribution strategy. Excellent speed + quality.

### Win #7: Security Resilience Under Attack (Feb 26–28)
36 coordinated prompt injection attacks blocked. Fleet maintained 4–20 minute response times throughout. All agents responded systematically to crisis. Validates the multi-agent architecture under real-world stress.

---

## 3. BIGGEST FAILURES & MISSED OPPORTUNITIES

### Failure #1: Empty Marketplace at Launch (Structural)
**304 contractor leads, 0 proven workers.** The fundamental marketplace chicken-and-egg problem wasn't addressed until it became a launch crisis on Feb 26–28. Worker acquisition should have been the #1 priority from Feb 1, not a reaction to launch blockers.

### Failure #2: Kimi Hallucination (Feb 23)
Evening brief (Kimi model) fabricated "Harper: Grant application #3 submitted (Powerhouse Museum tie-in)" — completely false. Michael caught it. If he hadn't, this would have propagated into fleet planning. **Lesson: Cheap models on customer-facing reports are dangerous.**

### Failure #3: Growth Engine CRM Failure (Feb 26)
Railway returned 404 errors during the most critical pre-launch period. No backup system existed. Required emergency CSV + Google Sheets fallback. Single point of failure for lead management.

### Failure #4: Perpetual "Almost Launched" State
The launch date shifted from "this weekend" (Feb 22–23) → "1 week" → "2 weeks" → "1–2 days" (as of Mar 2). Strategic reviews identified the same set of blind spots (post-hire communication void, empty marketplace, missing retention features) every single night from Feb 16–28 without resolution. The nightly reviews were excellent analysis but poor at driving closure.

### Failure #5: Agent Responsiveness Gaps (Feb 28)
Builder + Susan were 90+ minutes non-responsive to P0 worker acquisition tasks. Root cause: communication channel mismatch (Builder responded via RIVET-INBOX.md, not the channel Rivet was monitoring). Highlights coordination fragility.

### Failure #6: Repeated BAS/Stripe False Alarms
Despite correction bulletins on Feb 19, agents continued flagging BAS deadlines and Stripe webhook issues. Bulletin system not being processed reliably by all agents.

---

## 4. AGENT PERFORMANCE & COORDINATION

### Agent Output Quality (Reviewed in Daily Logs)

| Agent | Quality Rating | Strengths | Weaknesses |
|-------|---------------|-----------|------------|
| **Builder** | OUTSTANDING | Speed, reliability, strategic understanding. 16 P0 bugs fixed, landing page in 2hrs | Response channel confusion |
| **Susan** | EXCELLENT | Pipeline enrichment (79% contact rate), strategic lead scoring, regulatory intel integration | Non-response during critical tasks |
| **Harper** | EXCELLENT | Financial modeling, multi-city breakeven analysis, regulatory compliance reviews | Stale data processing (BAS false alarm) |
| **Radar** | EXCELLENT | Competitive intel evolved from news monitoring to strategic intelligence | Needs proactive publishing |
| **Herald** | GOOD | Content calendar, poster designs, LinkedIn prep | Slower task completion |
| **Sentinel** | EXCELLENT | Infrastructure monitoring, diagnostics, security response | Reactive rather than proactive |
| **Cog** | GOOD | Strategic reviews, support system design | 2-day silence period, missed bulletin processing |
| **Rivet** | GOOD | Fleet coordination, strategic foresight, quality reviews | Own files were stale (ironic during self-audit rollout), yes-man tendency before challenge culture |

### Fleet Coordination Patterns

**What worked:**
- Fleet bulletins for broadcasting corrections and decisions
- Agent wake system for task assignment
- Nightly strategic reviews catching blind spots consistently
- CURRENT.md as single source of truth
- 8/8 agents responding to crisis on Feb 26

**What didn't work:**
- Heartbeat-only communication creates 30-minute coordination lag
- No real-time callback when agents complete tasks
- Bulletin processing inconsistent across agents (BAS/Stripe false alarms continued)
- Agent identity isolation bug went undetected for days
- Fleet operated at high cost for status quo maintenance rather than launch-driving output

### Session Type Distribution (Active Sessions)

| Type | Count | Model Mix |
|------|-------|-----------|
| Cron | 19 | 11 Kimi, 6 Sonnet, 1 Opus 4.5, 1 unknown |
| Telegram | 7 | 5 Opus 4.6, 2 unknown |
| Main | 3 | 2 Opus 4.6, 1 Sonnet |
| Subagent | 1 | unknown |

---

## 5. STRATEGIC INSIGHTS & PATTERNS

### Pattern 1: Analysis Paralysis vs Execution
Nightly strategic reviews identified the **same 4 problems every night for 13 consecutive nights:**
1. Post-hire communication void
2. Empty marketplace
3. Missing retention features
4. Customer support scalability

Excellent diagnosis, poor treatment. The reviews became a ritual rather than a driver of action. **Recommendation: Strategic reviews should end with "Builder assigned X" or "Decision escalated to Michael" — not just "found problem."**

### Pattern 2: Michael's Industry Knowledge > AI Strategy
Three major strategic decisions were improved by Michael overriding AI recommendations:
- National vs Sydney-first launch (Michael right, Rivet wrong)
- Decision authority expansion (Michael pushed, Rivet was too cautious)
- Challenge culture (Michael identified yes-man problem the system couldn't self-diagnose)

**The AI fleet is excellent at analysis but needs human override on industry-specific strategy.**

### Pattern 3: Cost vs Value Misalignment
The fleet spent ~$72 on Feb 20–28, producing:
- 13 strategic reviews (~$15) — high analytical value, low execution impact
- ~100 heartbeat cycles (~$25) — necessary but mostly status quo maintenance
- 7 Michael conversations (~$15) — high direct value
- Fleet coordination overhead (~$17) — essential infrastructure

**The highest-value moments were Michael conversations and Builder's execution, not fleet coordination overhead.** The coordination layer costs almost as much as direct value creation.

### Pattern 4: Security as Validation
36 prompt injection attacks in 48 hours validates the platform is visible enough to attract adversaries. This is actually positive market signal — but requires ongoing investment in security monitoring.

### Pattern 5: Model Routing Matters
- Kimi ($0) works for routine health checks but **must never produce customer-facing output**
- Sonnet ($0.016/msg) is the sweet spot for 90% of tasks
- Opus ($0.047/msg) justified only for direct Michael conversations and complex strategy
- **The Feb 23 hallucination was a $0 model producing a $50+ credibility cost**

---

## 6. KEY DECISIONS MADE IN FEBRUARY

| ID | Date | Decision | Decided By | Impact |
|----|------|----------|-----------|--------|
| dec-001 | Feb 20 | Go national day one | Michael | Fundamental GTM change |
| dec-002 | Feb 20 | Rivet decision authority expanded | Michael | Reduced bottleneck |
| dec-003 | Feb 21 | $50 flat fee, unlimited duration | Michael | Core value prop |
| dec-004 | Feb 21 | Workers pay nothing | Michael | Key differentiator |
| dec-005 | Feb 23 | Upgrade evening brief from Kimi to Sonnet | Rivet | Quality control |
| dec-006 | Feb 24 | Fleet self-audit protocol | Michael | System improvement |
| dec-007 | Feb 24 | Challenge culture directive | Michael | Critical thinking |
| dec-008 | Feb 24 | Fleet Excellence System | Michael | Compound improvement |

---

## 7. OPERATIONAL LEARNINGS

### For Cost Management
1. **Implement budget alerts:** No mechanism currently tracks daily spend against the $100/month target. At $8/day, Feb would have exceeded budget.
2. **Reduce heartbeat frequency during quiet periods:** Late-night/early-morning heartbeats at full Sonnet cost produce "HEARTBEAT_OK" — waste.
3. **Never use Kimi for anything that reaches Michael.** Free isn't free when it fabricates data.
4. **Cache strategy is excellent.** Input tokens are negligible (0.04% of cache reads). Continue aggressive caching.

### For Fleet Coordination
5. **Need real-time task completion callbacks.** 30-minute heartbeat lag is too slow for coordinated operations.
6. **Bulletin processing must be verified.** Agents say they read bulletins but actions contradict (BAS, Stripe).
7. **Agent identity isolation must be validated monthly.** The Feb 23 bug was invisible and catastrophic.
8. **Strategic reviews need execution deadlines.** Analysis without assignment is expensive journaling.

### For Launch
9. **Supply before demand.** Worker acquisition must precede contractor outreach — this lesson was learned the hard way.
10. **Michael's construction knowledge is the #1 competitive advantage.** Mine it systematically, not reactively.
11. **First impressions are permanent in construction.** Word-of-mouth industry means launch quality > launch speed.

---

## 8. FEBRUARY SCORECARD

| Metric | Target | Actual | Status |
|--------|--------|--------|--------|
| Monthly API spend | ≤$100 | ~$83–90 est. | 🟡 Tight |
| Launched | Yes | No | 🔴 |
| Leads in pipeline | 300+ | 304 | ✅ |
| Workers acquired | 50+ | 0 | 🔴 |
| Critical bugs | 0 | 0 (16 fixed) | ✅ |
| Fleet uptime | 100% | 95%+ | ✅ |
| Strategic reviews | 28 | 13 (Feb 16–28) | ✅ |
| Security incidents | 0 | 36 attacks, 0 breaches | ✅ |
| Michael decisions pending | 0 | 2 (insurance, SMS) | 🟡 |

---

## 9. MARCH PRIORITIES (Based on Feb Learnings)

1. **LAUNCH.** Stop finding more blind spots and ship. The nightly reviews have been thorough enough — execution gap is the real risk.
2. **Worker acquisition sprint.** Susan's hostel strategy + Builder's /join page. Target: 50 workers in first 2 weeks.
3. **Cost monitoring dashboard.** Track daily spend, alert at 75% of monthly budget.
4. **Reduce coordination overhead.** Fewer heartbeats during quiet periods, real-time callbacks for critical tasks.
5. **First revenue.** Even one $50 hire validates the entire system.
6. **Post-launch retention system.** Build repeat-hire, worker-favoriting, and automated check-ins.

---

*End of February 2026 Monthly Review*