SaaStr Runs 30 AI Agents With 3 Humans. The Problems Nobody Talks About Are All Operational.

Everyone's talking about building agents. Almost nobody's talking about managing them.

SaaStr published the most granular operational report I've seen on running a multi-vendor AI agent fleet: 30 agents, 3 humans, across Agentforce, Artisan, Qualified, Monaco, and custom builds. Five issues stood out, and every one of them surprised me.

First: you need daily 1-on-1s with each agent. Not weekly. Daily. The agents drift, hallucinate edge cases, and develop behavioral patterns that compound if you don't catch them within 24 hours. This alone destroys the "set it and forget it" fantasy.

Second: no orchestration product exists. SaaStr runs four AI sales agents simultaneously. Each specializes in a different part of the funnel. Nothing coordinates them. Not MCP. Not any API. Not any product on the market. They built custom glue. This is the biggest gap in the agent ecosystem right now and I don't see anyone close to solving it.

Third: the 90/10 buy vs. build rule. Only vibe-code an agent when nothing commercial exists. The maintenance burden of custom agents is brutal, and commercial vendors iterate faster than your internal team can.

Fourth: Marketo is dying as a data source. Legacy marketing automation platforms are atrophying because agent data flows don't map to their schemas. If your agents depend on traditional MAP data, start planning the migration.

Fifth, and this is the one that'll bite you: vendors ship breaking changes without migration paths. Your downstream agents break silently. No changelog. No deprecation warning. Just broken pipelines at 2am.

The takeaway for builders: deploying agents is the easy part. Managing a fleet of agents from multiple vendors, keeping them coordinated, catching drift, surviving vendor updates. That's where the actual work lives. If you're planning to scale past five agents, budget 3x the ops time you think you'll need.