Steve Yegge gastou milhares/mês em agents. Matou tudo.

Steve Yegge gastou milhares/mês em agents. Matou tudo.

Você é founder de SaaS.

Você lê notícia:

• "Steve Yegge (famoso tech leader) shut down projeto AI agents (Gas Town)"
• "Reason: Agents não são confiáveis (don't complete tasks)"
• "Cost: Gasto milhares de dólares/mês em subscriptions"
• "Result: Projeto morreu. Investimento perdido."
• Your reaction: "Wait. Steve Yegge (super inteligente) tentou e falhou?"
• Your anxiety: "Se Yegge não conseguiu, como EU consigo?"
• Real question: "Estou também gastando dinheiro em agents que não funcionam?"
• Bigger realization: "Se agentes não funcionam pra coding (task mais clara), como funcionarão pra support/sales (ambíguo)?"

Seu problema AGORA:

Você tem SaaS com agentes:

• Monthly token cost: R$ 10K-50K (agents rodam 24/7)
• Your assumption: "Agents vão reduzir custo de headcount (suporte, vendas)"
• Reality: "Agents estão quebrando, customers reclamando, ROI é negativo"
• Your question: "Como sei se agentes estão gerando lucro ou só queimando dinheiro?"
• Answer from Yegge: "Agents são unreliable. Task completion é muito baixo. Vale a pena? Questionável."
• Your realization: "Estou em situação similar. Gastando muito, retorno baixo."

---

O Caso Steve Yegge: Como Hype Vira Realidade

Por que até especialistas falham com agents

=== WHO IS STEVE YEGGE? ===

Background: ├─ Ex-Google (Senior Engineer) ├─ Ex-Amazon (VP of Platform Engineering) ├─ Famous in tech (wrote "Platforms > Products") ├─ Expertise: System design, infrastructure, coding ├─ Intelligence: Top 1% (literally) └─ Resources: Capital to invest in R&D

His AI Agent Project (Gas Town): ├─ Vision: "AI agents for coding (reduce developer work)" ├─ Investment: "Thousands of dollars per month" ├─ Duration: ~12+ months (based on timeline) ├─ Tech stack: "Top AI coding agents (Copilot, custom agents)" ├─ Outcome: "SHUT DOWN (failed)" └─ His admission: "Agents are unreliable. Task completion is too low."

=== THE PAINFUL TRUTH ===

Yegge's findings: ├─ Agents: Look promising (in demos) ├─ Reality: "Can't reliably complete tasks" ├─ Cost: "Many thousands per month (expensive)" ├─ ROI: "Built only 1 real project (Gas Town) with agents" ├─ Lesson: "Agents ≠ productivity gain (yet)" └─ Implication: "If Yegge (top 1% engineer) can't make agents work, who can?"

=== THE PATTERN ===

What we're seeing: ├─ 2023-2024: "AI agents are the future! (Hype)" ├─ Mid-2024: Companies throw millions at agents ├─ Late 2024: "Wait, agents don't actually work (Reality)" ├─ Now: High-profile projects shutting down (Yegge, others) ├─ Pattern: "Hype cycle = deploy agents, discover they fail, kill projects" └─ Question: "Is this your timeline too?"

=== THE GAP ===

Why agents fail in practice: ├─ Gap 1: Demo vs Real-world │ ├─ Demo: Agent does task perfectly (curated example) │ ├─ Real-world: Agent fails on variations (poor generalization) │ └─ Result: Demo looks great, actual use sucks ├─ Gap 2: Controlled vs Chaotic │ ├─ Controlled: Coding task (clear goal, clear success metrics) │ ├─ Chaotic: Support/sales (ambiguous goals, unclear success metrics) │ ├─ Agents work ~60% in controlled (like coding) │ ├─ Agents work ~20% in chaotic (like support) │ └─ Result: Even in best case, agents fail 40% of time ├─ Gap 3: Single task vs Continuous │ ├─ Single task: Complete 1 request (achievable) │ ├─ Continuous: Handle 1000 requests/day (reliability 99.9% required) │ ├─ At 60% success rate: 400 failures/day (unacceptable) │ └─ Result: Agents can't scale to production requirements └─ Gap 4: Cost vs Benefit ├─ Cost: R$ 10K-50K/month (token usage) ├─ Benefit: Reduce 0.5 FTE headcount (R$ 60K/year = R$ 5K/month value) ├─ Math: Cost (R$ 10K) > Benefit (R$ 5K) = Negative ROI └─ Result: Agents are expensive way to replace cheap labor

---

A Realidade: Agent Economics não batem

Por que seu SaaS com agentes pode estar no vermelho

=== COST STRUCTURE ===

Your agent infrastructure (monthly): ├─ Token costs: R$ 5K-20K (depends on volume) ├─ API subscriptions: R$ 1K-5K (multiple agent services) ├─ Infrastructure: R$ 2K-10K (servers, databases) ├─ ML engineer: R$ 20K (to maintain/improve agents) ├─ Monitoring/debugging: R$ 2K (to catch failures) ├─ Total monthly: R$ 30K-50K ├─ Total annual: R$ 360K-600K └─ Per agent: R$ 15K-30K/month per 1 agent

=== BENEFIT STRUCTURE ===

What you're replacing (monthly): ├─ Option A: Support agent (cost to replace with AI agent) │ ├─ Human support agent salary: R$ 3K-5K/month │ ├─ AI agent cost: R$ 15K-30K/month (per agent) │ ├─ Savings: -R$ 10K-25K/month (you're losing money) │ └─ Math: Human is 3-6x cheaper than AI agent ├─ Option B: Sales development rep (cost to replace) │ ├─ Human SDR salary: R$ 5K-10K/month │ ├─ AI agent cost: R$ 15K-30K/month │ ├─ Savings: -R$ 5K-25K/month (you're losing money) │ └─ Math: Human is 2-6x cheaper than AI agent └─ Option C: Reduce response time (soft benefit) ├─ Benefit: "Faster responses to customers" ├─ Monetization: None (can't charge more for faster) ├─ Value: R$ 0/month (soft benefit, not hard revenue) └─ Math: Soft benefits don't offset hard costs

=== THE UNIT ECONOMICS PROBLEM ===

Simple math: ├─ Cost per month: R$ 30K-50K ├─ Benefit per month: R$ 0-10K (replacing 1 FTE) ├─ Margin: -R$ 20K-50K/month (NEGATIVE) ├─ Annual burn: R$ 240K-600K/year ├─ Payback period: NEVER (you're losing money forever) └─ Question: "Why are you doing this?"

=== THE RELIABILITY PROBLEM ===

Yegge's core finding: ├─ Agent task completion rate: ~60% (industry average) ├─ Success rate needed for support: 99%+ (customer expectation) ├─ Gap: 39% of tasks fail (unacceptable) ├─ Implication: "Agent can't handle 39% of requests" ├─ Fallback: "Human has to handle failed agent tasks" ├─ Result: "You're paying for agent + human (double cost)" └─ Real cost: R$ 30K (agent) + R$ 3K (human backup) = R$ 33K/month

=== THE ADOPTION PROBLEM ===

Customers don't trust agents: ├─ Customer psychology: "Agent = automation = lazy = bad service" ├─ Reality: "Agents actually provide faster service (if they work)" ├─ But customers feel: "I'm talking to a bot (impersonal)" ├─ Result: "Agents lower satisfaction even if faster" ├─ Data: Companies report 10-20% NPS drop after deploying agents ├─ Translation: "You saved money on cost but lost money on churn" └─ Math: Churn loss (R$ 100K) > Cost savings (R$ 10K) = Bad trade

=== THE YEGGE SIGNAL ===

What his shutdown means: ├─ Signal 1: "Agents are not ready for production (yet)" ├─ Signal 2: "Economics don't work (cost > benefit)" ├─ Signal 3: "Reliability is the bottleneck (60% completion is too low)" ├─ Signal 4: "Hype is ahead of reality (30% ahead)" ├─ Signal 5: "Winners will be those who wait, then implement properly" └─ Implication: "Deploying agents NOW = burning money"

---

A Solução: ROI-Driven Agent Strategy

Como fazer agentes funcionarem (economicamente)

=== PHASE 1: AUDIT YOUR CURRENT STATE ===

[ ] Current agent investment: ├─ Monthly token cost: R$ ___ ├─ Monthly API subscriptions: R$ ___ ├─ Monthly infrastructure: R$ ___ ├─ Monthly ML engineer: R$ ___ ├─ Total monthly: R$ ___ └─ Total annual: R$ ___

[ ] Current agent performance: ├─ Task completion rate: ___% (goal: >90%) ├─ Customer satisfaction: ___% (goal: >80%) ├─ Time to resolution: ___ minutes (goal: <10 min) ├─ Cost per resolved task: R$ ___ (goal: <R$ 5) └─ Is agent saving money? YES/NO

[ ] Current ROI: ├─ Cost: R$ ___ /month ├─ Benefit: R$ ___ /month (cost reduction) ├─ Net: R$ ___ /month (positive or negative?) ├─ Payback period: ___ months (or never) └─ Should you keep running? YES/NO

=== PHASE 2: DEFINE PROFITABLE USE CASE ===

Not all tasks are equal (for agents): ├─ High-ROI tasks (good for agents): │ ├─ FAQs (predictable, rules-based) │ ├─ Intake forms (collect info, structured) │ ├─ Ticket categorization (classify request) │ ├─ Simple refunds (clear policy) │ ├─ Escalation routing (if-then logic) │ ├─ Success rate: 85-95% (agent can handle) │ ├─ Cost to replace: R$ 2K-5K/month (1 FTE) │ ├─ Agent cost: R$ 3K-8K/month │ ├─ ROI: Marginal (break-even or slightly negative) │ └─ Recommendation: Deploy agents here ├─ Medium-ROI tasks (maybe): │ ├─ Troubleshooting (multiple paths) │ ├─ Feature requests (creative writing) │ ├─ Account questions (context-dependent) │ ├─ Success rate: 60-75% (agent struggles) │ ├─ Cost to replace: R$ 5K-10K/month │ ├─ Agent cost: R$ 10K-20K/month │ ├─ ROI: Negative (don't deploy) │ └─ Recommendation: Wait for better models └─ Low-ROI tasks (don't use agents): ├─ Complex issues (need human judgment) ├─ Emotional support (empathy required) ├─ Product feedback (strategic decisions) ├─ Success rate: 20-40% (agent fails) ├─ Cost to replace: R$ 10K+/month ├─ Agent cost: R$ 20K+/month ├─ ROI: Negative (don't deploy) └─ Recommendation: Use humans only

=== PHASE 3: CALCULATE TRUE ROI ===

Real example (Scenario A: Break-even): ├─ Use case: FAQ automation (high-ROI task) ├─ Volume: 100 customer questions/day ├─ Agent success rate: 90% (good, for this task) ├─ Agent handles: 90 questions/day ├─ Human handles: 10 questions/day (failures + complex) ├─ Agent cost: R$ 5K/month (token + infra) ├─ Human cost (partial): R$ 1K/month (20% of 1 FTE) ├─ Cost savings: R$ 3K/month (vs 100% human-based) ├─ Net ROI: -R$ 2K/month (still negative, but close) ├─ Timeline: In 6 months, improvements could flip to positive └─ Decision: Deploy, but monitor closely

Real example (Scenario B: Profitable): ├─ Use case: Ticket intake form (very high-ROI) ├─ Volume: 500 tickets/day ├─ Agent success rate: 95% (excellent, for structured task) ├─ Agent handles: 475 tickets/day (collects info, routes) ├─ Human handles: 25 tickets/day (only ambiguous ones) ├─ Agent cost: R$ 3K/month ├─ Human cost (partial): R$ 500/month (5% of 1 FTE) ├─ Cost savings: R$ 4.5K/month (vs having 2 FTE intake staff) ├─ Net ROI: +R$ 1.5K/month (PROFITABLE) ├─ Annual ROI: +R$ 18K └─ Decision: Deploy now (clear positive ROI)

Real example (Scenario C: Disaster): ├─ Use case: Complex support (low-ROI task) ├─ Volume: 50 complex issues/day ├─ Agent success rate: 40% (terrible) ├─ Agent handles: 20 issues/day (successfully) ├─ Human required for: 30 issues/day (agent failed) ├─ Agent cost: R$ 15K/month ├─ Human cost (full): R$ 15K/month (2 FTE to handle all) ├─ Cost savings: R$ 0 (no savings, agent doesn't help) ├─ Net ROI: -R$ 15K/month (DISASTER) ├─ Annual cost: -R$ 180K/year └─ Decision: Kill this use case (Don't deploy)

=== PHASE 4: BUILD MEASUREMENT FRAMEWORK ===

You MUST track: ├─ [ ] Task completion rate per task type │ ├─ Target: 85%+ for deployment │ ├─ Frequency: Weekly │ └─ Alert: If drops below 80%, investigate ├─ [ ] Customer satisfaction (CSAT) per agent response │ ├─ Target: 75%+ (can drop 10% from human baseline) │ ├─ Frequency: Weekly │ └─ Alert: If drops below 70%, kill use case ├─ [ ] Cost per resolved task │ ├─ Formula: (Token cost + infra) / (Successfully completed tasks) │ ├─ Target: 30-50% cheaper than human baseline │ ├─ Frequency: Weekly │ └─ Alert: If cost goes up, ROI is negative ├─ [ ] Churn impact (customers leaving after agent interaction) │ ├─ Target: 0% increase in churn │ ├─ Frequency: Monthly │ └─ Alert: If churn increases >2%, too many agent failures └─ [ ] Financial ROI ├─ Formula: (Cost savings - Agent cost) / Agent cost ├─ Target: 10%+ monthly ROI (to justify cost) ├─ Frequency: Monthly └─ Alert: If ROI negative for 2+ months, kill use case

=== PHASE 5: DECIDE (Kill or Keep) ===

Decision matrix:

┌──────────────────┬──────────────┬────────────────────────────────────┐ │ Metric │ Good ✓ │ Action │ ├──────────────────┼──────────────┼────────────────────────────────────┤ │ Completion │ >85% │ Keep │ │ rate │ 70-85% │ Optimize (retraining, prompts) │ │ │ <70% │ Kill use case │ ├──────────────────┼──────────────┼────────────────────────────────────┤ │ CSAT │ >75% │ Keep │ │ │ 60-75% │ Monitor, optimize │ │ │ <60% │ Kill (customers hate agents) │ ├──────────────────┼──────────────┼────────────────────────────────────┤ │ Cost/task │ <R$ 3 │ Very profitable, scale up │ │ │ R$ 3-10 │ Profitable, keep │ │ │ R$ 10-20 │ Break-even or negative, optimize │ │ │ >R$ 20 │ Kill (too expensive) │ ├──────────────────┼──────────────┼────────────────────────────────────┤ │ Churn impact │ -0% │ Good (no negative impact) │ │ │ +0-2% │ Acceptable (monitor) │ │ │ >+2% │ Customers hate agents, kill │ ├──────────────────┼──────────────┼────────────────────────────────────┤ │ ROI │ >0% (positive)│ Keep and monitor │ │ │ -50% to 0% │ Marginal, optimize or kill │ │ │ <-50% │ Kill immediately (burning cash) │ └──────────────────┴──────────────┴────────────────────────────────────┘

=== PHASE 6: THE YEGGE DECISION TREE ===

If your agent metrics are: ├─ All green (✓): Keep deploying, scale up ├─ Some red (✗): Kill failing use cases, keep winning ones ├─ All red (✗✗✗): Kill agents entirely (Yegge's situation) └─ Mostly red with 1-2 green: Narrow focus on winners

Special case (Yegge's situation): ├─ He spent: Many thousands/month ├─ He got: 1 real project (Gas Town) ├─ His ROI: Negative (project is dead) ├─ His decision: Kill agents ├─ Your lesson: Don't be like Yegge (broad approach, no selectivity) ├─ Better approach: Pick 1-2 high-ROI use cases, ignore rest └─ Result: Agents become profitable (on specific tasks)

---

Your Checklist: Are you like Yegge?

Signs you're burning money on agents (like he was)

=== RED FLAGS ===

[ ] You're running agents on ALL support tasks └─ Sign: You haven't narrowed down to high-ROI use cases └─ Risk: 70%+ of your token cost is wasted └─ Action: Kill agents on low-ROI tasks NOW

[ ] Your agent costs more than human baseline └─ Sign: R$ 20K/month on agents vs R$ 10K/month on human └─ Risk: You're paying 2x for worse service └─ Action: Reduce scope or kill agents

[ ] You don't track ROI metrics └─ Sign: You don't measure completion rate, CSAT, cost/task └─ Risk: You don't know if you're profitable (probably not) └─ Action: Build measurement framework TODAY

[ ] Your customers prefer humans (after trying agents) └─ Sign: CSAT drops >10% after agent deployment └─ Risk: Churn is increasing (hidden cost) └─ Action: Kill agents immediately

[ ] You have multiple agents doing same task └─ Sign: You're duplicating work across 3-5 agent services └─ Risk: Massive token waste └─ Action: Pick ONE best agent, kill others

[ ] Your agents are unreliable (completion <70%) └─ Sign: Humans have to redo agent work └─ Risk: You're paying 2x (agent + human) └─ Action: Pause agents until completion >85%

[ ] You have dedicated ML engineer just for agents └─ Sign: R$ 20K/month for one person = very expensive └─ Risk: Cost justifies only if agents deliver 10x value └─ Action: If ROI negative, fire them (Yegge did)

[ ] Your agent project is 12+ months old with negative ROI └─ Sign: Same timeline as Yegge (sunk cost fallacy) └─ Risk: You're going to hit his decision (kill it) └─ Action: Kill NOW rather than after 18 months

=== IF YOU CHECKED 3+ BOXES ===

You are likely in Yegge's position (burning money): ├─ Action 1: Stop expanding agent use cases ├─ Action 2: Audit ROI on current agents ├─ Action 3: Kill agents on low-ROI tasks ├─ Action 4: Keep only 1-2 profitable use cases ├─ Action 5: Reduce infrastructure costs (kill unused agents) ├─ Result: Stop bleeding cash (like Yegge did) └─ Timeline: 30 days

=== IF YOU CHECKED 0 BOXES ===

You might be profitable: ├─ Action: Continue monitoring metrics weekly ├─ Action: Scale up successful use cases ├─ Action: Look for new high-ROI opportunities ├─ Result: Agents become strategic advantage └─ Timeline: Ongoing

---

Conclusão: The Yegge Reality Check

O que sua situação é:

1. Você é provavelmente em situação similar a Yegge (not as smart, but same economics)

   • You think: "Agents will save us money."
   • Reality: "Agents cost more than they save (if you measure properly)."
   • Implication: "Your agent investment is likely negative ROI."

2. Hype is ahead of reality (agents aren't ready yet)

   • You think: "Agents are production-ready."
   • Reality: "Agents are 60-70% reliable (need humans for 30-40%)."
   • Implication: "You're paying for agents + humans (double cost)."

3. Economics matter more than features (Yegge learned this hard way)

   • You think: "Better agent = win."
   • Reality: "Profitable agent = win (regardless of quality)."
   • Implication: "ROI is only metric that matters."

4. Selectivity beats broad deployment (Yegge's mistake)

   • You think: "Deploy agents everywhere."
   • Reality: "Deploy agents only on high-ROI tasks (FAQ, intake, routing)."
   • Implication: "Broad deployment = guaranteed negative ROI."

5. Time to decision matters (the longer you wait, the more you burn)

   • You think: "I'll give agents more time."
   • Reality: "Every month of negative ROI is money down the drain."
   • Implication: "Make kill/keep decision in 30 days (not 12 months)."

Your decision today:

• Audit ROI (do agents make money?)
• Kill low-ROI agents (stop bleeding cash)
• Focus on high-ROI use cases (FAQ, intake, routing)
• Build measurement framework (track completion, CSAT, cost)
• Check metrics weekly (not annually)
• Kill entire program if ROI negative for 2+ months (like Yegge did)

Na OpenClaw:

Ajudamos SaaS builders fazer agentes funcionarem economicamente:

• Agent ROI audit: Seu agente está lucrando ou queimando? (assessment)
• Use case selectivity: Identificar high-ROI tasks para agents (strategy)
• Reliability improvement: Aumentar completion rate (optimization)
• Cost reduction: Reduzir token spend sem perder qualidade (efficiency)
• Measurement framework: Track completion, CSAT, cost, churn (metrics)
• Kill/keep decision: Quando desabilitar agentes unprofitable (decision-making)

Você pode continuar gastando como Yegge (hoping it works eventually).

Ou você pode implementar ROI-driven strategy AGORA (stop bleeding, start profiting).

Agent ROI Audit | Use Case Selectivity | Profitability Framework →