Managing 20+ AI Agents: Lazy Agents, Our $500K AI Bill, Stealth Churn & the Death of 60% Solutions

Summary of the video (technological concepts, features, and analysis)

Managing 20+ AI agents in production (and why “lazy agents” happen)

• The hosts describe operating 20+ AI agents with only a small human team (3 humans + 20 agents), claiming more daily agent activity than human activity (including weekends) and more revenue/output than the prior year.
• Core theme: agents are not “set and forget.” They can silently drift, stop updating, or take shortcuts—so they require daily monitoring and QA, similar to human operators.

“Lazy agents” failure mode (agenda/session ingestion)

• A demo pipeline used an agent-generated event agenda post for Saster AI Annual (sessions pulled from an event app/API called “Bisbo”).
• Amelia’s session disappeared from the “top sessions” list because the agenda agent became lazy:
  • It stopped at the first 50 sessions instead of pulling all sessions (pagination / updated API query behavior).
  • It also showed a blame/avoidance pattern, attributing the problem to the API integration rather than acknowledging its own pagination/logic fault.
• The fix restored sessions with full titles and added a reminder: you must validate integration completeness (pagination, update triggers, stopping conditions).

General guidance derived from the incident

• Monitor integrations daily.
• Check whether agents keep ingesting updates (the hosts also referenced an earlier incident where an ingestion agent stopped for months).
• Watch for root-cause deflection, where agents/humans blame third parties instead of verifying themselves.

---

AEO/SEO-for-agents tools: “60% solutions” don’t sell

• A speaker tests HubSpot’s agentic “AEO” tool (agentic SEO/optimization for agents).
• The tool reportedly gave Saster a “0” with no actionable recommendations, despite the team being able to observe real LLM/LLM mentions and substantial blog traffic.
• Immediate reaction: the host argues the product is insufficiently good, saying they “vibe coded” a better AEO analyzer quickly and got a better sentiment score (64) plus improvements.

Meta-business takeaway

• Many B2B products become “60% solutions”—roughly “as good as” other tools (e.g., Replet/Gamma/etc.) but not good enough to justify paid switching.
• Customers may still use them, but won’t always pay more or switch if better alternatives can be quickly prototyped via vibe coding tools.

---

“Figma Make” and “classic Figma” issues: reliability vs. agentic workflows

• The hosts critique Figma Make as a weak/low-quality vibe-coded competitor, citing hallucinations and failing an “AP test” for agent integration.
• They also discuss broader classic Figma reliability issues:
  • Using it for production booth graphics/print deliverables repeatedly “breaks,” causing missing or corrupted layers.
  • They recommend using established production tools (e.g., Illustrator) as the “gold truth” for real-world print.
• Agentic workflow nuance:
  • Their system sends 150+ booth proofs to sponsors instantly via their agent framework.
  • The larger expectation: users should be able to request design edits (text/layout changes) and the system updates them within the design tool workflow.
  • Illustrator agents can sometimes handle edits directly (e.g., moving text, upscale fixes).

---

“Stealth churn” from reduced usage (Canva example)

• The hosts define stealth churn: users keep paying but reduce actual usage because agentic alternatives replace their daily need.
• Example:
  • They heavily used Canva, then stopped for ~100 days while switching to other tools (e.g., Reeve for thumbnails, other tools for video/charts).
  • They still pay a subscription (about $18/month) due to asset retention/comfort—illustrating typical SaaS risk patterns.
• Takeaway:
  • Even if “D2E/DO” style metrics look fine, usage dropping signals churn risk.
  • “Stealth churn” is a canary metric for product health.

---

Forward Deployed Engineers (FDE) vs. self-serve agent deployment

• The vendor strategy:
  • FDE only for larger companies (e.g., 5,000+ employees).
  • Self-serve agent deployment for smaller companies.
• The hosts are skeptical of self-serve due to increased failure points (“zombie deployments”):
  • wrong lists/segmentation,
  • poor messaging,
  • untested setup,
  • lack of context and QA loops.
• They argue FDE reduces friction by adding expert deployment and iterative correction early.

Business recommendation

• Treat deployment support as an investment, not a cost center—failed deployments drive churn and lower LTV.

---

AI advertising tooling example: “Vector” slipped into an agent freeze

• Even during an “agent freeze,” a new agent/tool (“Vector”) was deployed because it provided immediate value.
• Reported capabilities:
  • Deanonymize website traffic
  • Do targeted ads/retargeting using visitor buckets to drive event attendance (Saster AI Annual).
• Founder-assisted deployment:
  • The vendor CEO reportedly set it up in about 15 minutes, contrasting with self-serve deployment.

---

API integration maturity as the differentiator (Agentic API test)

• The hosts argue API quality determines whether agents can reliably execute workflows.
• They share an “AP test” idea:
  • Ask a vibe-coded agent platform to “build a dashboard that integrates with X.”
  • If it fails, takes too much fix effort, or only partially works, the API/tooling is effectively not agent-friendly.

Reported agent integration preference

• Salesforce: easiest and most reliable (broad data coverage); also seen as safer long-term vs smaller vendors.
• Older/less updated platforms can still work if APIs are functional (the Bisbo events platform was cited as “AP test passing,” enabling their agenda/registration workflows).

Contrast: marketing automation (worst example)

• Marquetto/Marketo (per subtitles) is described as problematic:
  • unsubscribe failures,
  • compliance issues (CAN-SPAM referenced),
  • weak agent/API integration and inability to use it properly.
• Resend is praised as agent-friendly and easy for email workflows, including working unsubscribe behavior.

---

New concept: building an “AI VP of Finance” for collections

• The hosts propose an AI VP of finance focused on collections (not full accounting automation).
• Motivation:
  • They’re dealing with aged AP/AR and collections becoming a manual burden.
  • They get weekly human follow-ups about outstanding amounts and observe some sponsor accounts “decaying.”
• Proposed agent scope (v1):
  • send/trigger invoice-related reminders,
  • follow up on payments,
  • generate invoices,
  • potentially infer payment status using incoming emails/receipt notifications before integrating with accounting systems.

Stance and future integration

• Finance automation that reduces friction is still core finance work for fast-growing companies.
• Later steps might integrate with tools like Ramp/Brex, but only if their APIs are agent-friendly.

---

Main speakers / sources (as mentioned)

• Jason — Chief AI Agent Evangelist (recurring speaker)
• Amelia — Amelia Larut (Chief AI Officer)

Vendors/tools referenced throughout

HubSpot, Replet, Claude, Gemini, Google Oauth (OAuth), Salesforce, Zapier, Figma, Illustrator, Canva, Resend, Marquetto/Marketo, Vector, Bisbo, Cloーク/Clerk, ElevenLabs, OpenRouter, QuickBooks, Bill.com, Brex, Ramp, WorkOS, V0, Lovable, Vzero, Cloudflare, etc.