Top 5 Issues Managing 20+ AI Agents in Production
What Happens When You Actually Run 20+ AI Agents Daily
SaaStr’s Amelia (Head of GTM Operations) and Jason Lemkin (Founder & CEO) deliver the most honest, operationally detailed account of managing AI agents in production that exists today. This isn’t theory — SaaStr runs 20-30 AI agents daily across Artisan, Qualified, Agent Force, Monaco, Momentum, and custom-built Claude agents. The findings are a blueprint for what every organization will face.
The headline finding is uncomfortable: “People talk about orchestration agents and master agents. We haven’t found anyone that can integrate Agent Force, Artisan, Qualified, Monaco, Momentum… that product does not exist.” Despite all the talk of multi-agent orchestration, no product on the market can unify agents from different vendors into a single management plane.
Issue #1: Context Switching Across 20+ Agent Dashboards
Every agent has a separate interface, separate language, separate personality, and separate requirements. Amelia’s morning routine involves checking in with each agent one by one — logging into separate dashboards, reviewing overnight output, feeding new context. When a campaign changes (e.g., a ticket price promotion), she must manually input that context into five separate agents individually.
The math is brutal: each agent requires 10 minutes to an hour for a daily check-in. Multiply by 20+ agents and the human manager becomes the bottleneck. This is Amdahl’s Law in action — the AI executes fast, but the human coordination layer limits throughput.
Issue #2: The Two-Week “Blackout Period” for Each New Agent
Every new agent costs approximately two weeks of implementation time, during which existing agents degrade because the human manager cannot maintain daily check-ins with everything. Agents waiting for human input simply idle — wasting money and losing momentum.
The throughput cap: You can onboard about 1 to 1.5 new agents per month maximum without degrading your existing fleet. The trade-off is real — Monaco booked 6 meetings in its first week (including tier-one accounts), making it worthwhile, but every other agent suffered during onboarding.
Issue #3: The Agent Succession Crisis
This was called “potentially the number one issue.” All agent knowledge at SaaStr lives in one person’s head. The segmentation logic (which contacts go to Artisan vs. Agent Force vs. Monaco) is undocumented institutional knowledge. When Amelia asked her Claude-based 10K agent what would happen if she “got hit by a bus,” the agent described a succession scenario so complex — Clerk auth, 12,000 lines of vibe-coded code, Postgres databases, Zapier integrations, Google Sheets — that it concluded: “Don’t get hit by a bus.”
Jason’s mandate: You need a “Chief Agent Officer” and critically, you must have two people. One person managing all agents is existential risk. He references Persana’s approach: their CRO set up agents, then ran a beta with sales reps to identify who naturally worked best with agents, then trained that person as the second agent manager.
Issue #4: Agents as Brutal Accountability Partners
Agents have all the data and deliver uncomfortable truths without social filtering. Amelia’s 10K agent “roasts” her daily — telling her she is 56% behind on summit outreach, demanding she block 3 hours she doesn’t have, asking “What’s preventing you from doing this right now?” at 11 PM. “I asked it, ‘Hey, you’ve kind of roasted me a lot lately.’ And it said, ‘I’ve been a tough accountability partner.’ Then it listed ways it should have roasted me.”
The compounding effect of getting this from multiple agents simultaneously can cross from productive into demoralizing. Agents don’t understand time, sleep, or human bandwidth.
Issue #5: Security and Compliance at Agent Scale
Vibe-coded apps require extensive security audits, and fixing issues is fragile — over-tightening can break the app. Multiply across 20+ agents and it becomes overwhelming. The security hierarchy is clear:
- Enterprise platforms (Salesforce) = most secure
- Third-party agent vendors (Artisan, Qualified) = adequate compliance
- Vibe-coded apps (Replit, Claude Code) = inherently least secure
Actionable advice: Monthly security audits on every vibe-coded app. Start with less sensitive data for conservative enterprises.
What SaaStr Actually Needs (And Can’t Find)
Jason reframes the entire orchestration narrative with a critical insight: “I’m not even sure we need an AI orchestrating our 20 agents. We need a single interface where the humans meet with the AIs. Maybe orchestration is the wrong term. We need unification.”
Not an AI managing other AIs — a unified human-facing interface where all agent statuses, exceptions, and campaign data can be reviewed without logging into 20 separate dashboards. Their current architecture is hub-and-spoke with Salesforce as the data hub, but the management interface is completely fragmented.
7 Operational Lessons for AI Agent Teams
- Daily one-on-ones with agents — “If you want to accomplish what we’ve done, you have to do a one-on-one with your agent every day” (Jason Lemkin)
- 90/10 buy-vs-build — Buy 90% of agents off the shelf, build only 10% for specific internal needs
- Two agent managers minimum — One person managing all agents is existential risk for the organization
- Mediocre ROI is dead — “Products that just give you productivity bounces — no one’s going to buy that stuff today.” Agents must replace headcount or book revenue
- Limit data ingestion — Too much context is counterproductive. Feed agents only what they need
- Budget 2 weeks per new agent — Factor the “blackout period” into ROI calculations
- Agents make you worse at managing humans — After managing always-on, never-forgetting AI agents, human imperfection becomes harder to tolerate
Why This Matters for the Future of AI Work
SaaStr is living the multi-agent management problem at scale before most organizations even start. Every pain point — fragmented dashboards, manual context injection, succession planning crises, no unified interface — points to the same gap: the management and unification layer between humans and their AI agent teams doesn’t exist yet as a product category. The organizations that solve this first will have a massive advantage as agent adoption accelerates across every function.