The AI Agent Broke Because We Locked It Down Too Hard
SaaStr's free pitch deck grader started failing hard during their 10,000-person conference. Of 305 submissions, 88 crashed outright. Of the 216 that ran, 53% returned an F grade. The problem was not bad decks. The problem was 14 guardrails.
Every time the agent pulled wrong data (projections instead of current ARR, TAM instead of actual revenue), Jason Lemkin added another rule. By guardrail 14, the agent was so paranoid it treated every deck with both current numbers and projections as suspect. That is every pitch deck ever written. Ambiguity defaulted to "no data," which stored as zero, which collapsed all sub-scores to zero, which output an F.
They scrapped the rules and rebuilt from scratch. The lesson: guardrails feel like free safety until they throttle the agent into doing nothing. Two guardrails are a feature. Fourteen are technical debt that breaks the product.
Same Spec, Different Platform, Different Agent
Amelia rebuilt their AI VP of Marketing (called 10K) on Lovable using the same spec that ran on Replit for months. Same data, same APIs, same instructions. The agents behave differently.
Replit's version returns three marketing ideas. Lovable's returns four. Lovable's is more aggressive: it recommended paid LinkedIn ads and a flash sale immediately. Replit's version has never suggested either. The Replit agent structures ideas like a B2C performance marketer. The Lovable one just says "here is the plan, go do it."
Two platforms, one spec, two personalities. The spec is not the agent. The platform shapes behavior as much as your instructions.
An Agent Now Handles Vendor Renewals
SaaStr's AI VP of Finance found a settings error they had missed for 8 years. Then an agent took over a vendor renewal negotiation. The vendor did not love it.
For sales teams deploying AI SDRs or outbound agents, the pattern is clear: building agents is easy now. Running them in production is where teams break. SaaStr went from -19% to +47% YoY revenue by shipping agents that actually work. The companies making this shift are not over-guardrailing, they are not under-guardrailing, and they are letting agents handle workflows humans used to own.
If you are evaluating AI SDR tools or sales automation agents, the question is not "can it send emails." The question is "what breaks when it runs for 90 days straight." SaaStr's answer: usually the guardrails you added to keep it safe.