Nearly 30 AI Agents in Production: The 5 Real Problems Nobody Talks About

Anyone who has actually put AI agents into production knows the landscape changes dramatically once the initial excitement wears off. The theory looks great, the demos are impressive, but the day-to-day reality of dozens of agents running simultaneously reveals a world that no sales pitch ever shows you.

SaaStr reached nearly 30 agents in production over about 10 months, operating across different areas of the business — from outbound sales to inbound lead qualification and internal operations. What started as a couple of experiments turned into a complex infrastructure of agents and applications running throughout the entire go-to-market stack. And what they discovered along the way never shows up in any demo, webinar, or polished case study.

The most surprising takeaway? Managing 30 AI agents is harder than managing the 12 humans the company had at peak headcount. Not harder in every way, but harder in ways nobody anticipated. The real challenges aren’t what most people imagine — it’s not about technology failing or models lacking accuracy. It’s about agent management, human processes, and organizational decisions that nobody prepared anyone to make.

Data and integration security, onboarding new agents without blowing up the current operation, unifying contexts that don’t talk to each other — these are the problems that surface after you’re already in the thick of it. 🏃 Below are the 5 main problems that came up in practice, with real examples, plus a bonus that might be the most uncomfortable one of all.

1. The Brutal Cost of Context Switching Between Agents

When you have one or two agents, context is simple to manage. When you have 20, 25, nearly 30 agents running in parallel — each with its own context window, its own data, its own integrations, and crucially, its own interface — things start falling apart in ways that aren’t obvious at all.

At SaaStr, they stopped thinking of these agents as tools. They started seeing them as 20 different AI employees, each with a distinct personality, different needs, and a separate dashboard that someone has to access every single day. Some send data back to Salesforce. Others don’t. Some run on Claude. Others don’t. They all ingest context in similar ways, but different enough that switching between them demands real mental effort.

The team’s morning routine illustrates the problem well: the day starts with a deep session with 10K, an internal agent that acts as VP of Marketing and runs on Claude and Replit. It literally tells you what to do that day — tickets, sponsors, outreach, campaigns. Then the team moves on to the external sales agents: Artisan, Qualified, AgentForce, and Monaco. That’s four separate dashboards, four different interfaces, four agents that need individual human review.

And here’s the core problem: these agents don’t talk to each other.

When SaaStr ran a ticket pricing promotion for SaaStr AI Annual, the team had to manually update five different agents with the same context. Artisan needed to know. Qualified needed to know. AgentForce needed to know. 10K already knew because it had created the promotion — but it was pushing for an immediate LinkedIn ad launch while the team was still informing the other agents about the price change.

A lot of people talk about orchestration agents and master agents. SaaStr hasn’t found a single one that works in practice. Despite everything out there — MCP, APIs, and the like — there is no product today that can integrate AgentForce, Artisan, Qualified, Monaco, and custom-built tools into a single management layer. That product simply didn’t exist in early 2026.

What the team actually needs isn’t orchestration — it’s unification. A single interface where humans meet the AIs. Maybe with some automation layered on top. But the agents are already running on their own. The bottleneck is the human side.

The practical lesson: you’re going to have a one-on-one with every agent every day. Not weekly. Daily. If you wait a week, the volume of output is so high that everything will be outdated by the time you come back. And if you skip the daily check-in, you’re throwing money away — because most of these agents are sitting there waiting for your inputs. Without them, they’re idle.

2. The Blackout Period When Adding a New Agent

Onboarding a new AI agent into an operation that’s already running is way more delicate than it seems. At SaaStr, every new agent costs at least two weeks of ramp-up time. They managed to cut that down from the month-plus it took in the beginning, but two weeks remains the floor — even with excellent vendor support.

And during those two weeks, your existing agents degrade.

When they were onboarding Monaco, a new SDR agent, the team couldn’t dedicate the usual time to their other agents. Some literally sat idle because nobody provided new contact lists or updated their campaigns. An outbound agent that has already worked through its entire contact list and is waiting for new names? It’s doing exactly nothing. Zero output. You’re paying for it and getting nothing in return.

Monaco was operational in about a week and a half. In its first week running, it reached out to 64 people and booked 6 meetings, including tier-one accounts. Was the trade-off worth it? Yes. But it needs to be planned ahead of time.

The math SaaStr landed on is this: one to one-and-a-half new agents per month, max. More than that and you start stalling — you can’t take care of your current agents while integrating new ones. Before adding another agent, the right question is: can I absorb a two-week blackout period right now? If you plan for it, it works. If you try to squeeze it in thinking you can pull it off in a day, it won’t. 😬

3. The Agent Succession Planning Crisis

This might be the biggest problem on the entire list.

At SaaStr, all the knowledge about how agents are segmented — which contacts go to Qualified, which to Artisan, which to Monaco, which to AgentForce — lives inside a single person’s head. If that person disappears from the operation for any reason, the agents effectively stop functioning in a coordinated way.

The team ran a revealing test: they asked the agents themselves what they’d do if the Chief AI Officer suddenly vanished. The answers were… eye-opening.

• The Claude version running 10K said it would need certain documents handed over — documents stored locally on the operator’s laptop and probably nowhere else. It listed the campaigns planned for March and April. And then, unexpectedly, it flagged what it called the vibe of SaaStr Annual — the agent had noticed that the phrase a moment in time appeared frequently in the copy and positioning, and it wanted whoever took over to understand that. The agent wanted to transfer the essence of the brand. Nobody saw that coming.
• The SaaStr Sponsors app — a tool built with vibe coding and 12,000 lines of code — was even scarier. When asked how someone new would get up to speed, it detailed the authorization system (each sponsor company uses Clerk, which requires a personal login), the manually created tier system, the hard-coded sponsor data, the Postgres database, file uploads going to Replit object storage, and submissions syncing with Google via Zapier. The handoff document read like a nightmare.

The agent’s final recommendation? Don’t get hit by a bus.

There is no agent today capable of managing all other GTM agents on its own. Not right now, in 2026. And if you’re a bigger company with one person who understands the agents and that person leaves, you’re facing existential risk.

What to do: the moment you find someone capable of deploying and managing agents competently, recruit a second person immediately. Divide and conquer. You need at least two people. And when it comes to hiring, the test is simple — give the candidate credits on Replit or Claude Code and tell them to build something that automates GTM. That’s it. That’s the interview.

4. The Agent as a Brutally Honest Mirror

SaaStr’s VP of Marketing agent, 10K, doesn’t spare anyone. Every single day.

You’re behind on summit outreach. You needed to have sent 200 emails by February 15. You sent 87. You’re 56% below target. Block 3 hours today to catch up.

Is it wrong? No. But when it’s 11 PM and the agent is asking what’s stopping you from launching the LinkedIn ads right now — and the honest answer is that you need to sleep — it starts feeling less like accountability and more like harassment.

When the team asked 10K to show examples of times it had been tough on them, the response was priceless: the agent said that, looking at the transcripts, it actually hadn’t been that tough — it had just been a firm accountability partner. And then it listed five ways it should have been even tougher.

The agent doesn’t care about your feelings. It has all the data. It knows your schedule. It knows your goals. It knows exactly where you’re falling short. And it will tell you, every day, without any of the social graces a human colleague would have.

When you multiply that by more than 20 agents — each one pointing from a different angle at where you’re falling behind — it can become genuinely demoralizing. For real. There are days when the SaaStr team has to tell the agent: take it easy today, please. You can handle hard truths most days, but not every day, from every agent, all the time.

The upside is powerful: these agents pursue objectives in a fundamentally different way than humans. A human on the marketing team might say they only want to work on ads and nothing else. That’s a personal goal. The agent pursues business goals with zero ego involved. It doesn’t understand why you can’t keep up — but it also doesn’t get distracted, doesn’t lose motivation, and doesn’t slack off.

5. Security and Compliance: The Problem Nobody Wants to Face Early

Security in multi-agent AI environments is a topic that tends to get pushed to the back burner — and that’s exactly why it becomes a nightmare later on. At SaaStr, every security audit on one of the apps built with vibe coding takes days to fix everything that comes up. And when you start implementing security fixes, the app becomes fragile. The agent overcorrects, locks everything down to the point of making the app unusable, and then you have to undo the fixes one by one.

The security hierarchy SaaStr identified works like this:

• Salesforce and enterprise platforms sit at the top. SOC 2 compliant, ISO certified, a decade of built-in data security. That’s one of the big reasons to use Salesforce as the central hub — security has already been thought through for you.
• Third-party agent startups sit one tier below. They’re startups. Inherently less secure than Salesforce. Less secure because they’re younger companies and because agents can do unpredictable things with your data — publish information externally, share it in contexts you didn’t anticipate.
• Apps built with vibe coding sit at the bottom. Claude Code, Replit, Lovable — all have added more security features. But anything custom-built is less secure than a mature third-party tool. All of SaaStr’s apps have customer data flowing through them. It’s not credit card data, but it’s real data, and you’re taking a calculated risk.

The bare minimum you should do: before launching any app built with vibe coding, and then at least once a month, have the agent run a deep security audit. Don’t assume your apps are secure by default. And for third-party tools, ask vendors directly: what’s your current compliance posture? What are you doing to keep data secure? You’d be surprised how many people never ask. 🔒

Bonus: Managing Agents Is Making Us Worse at Managing Humans

This one is slightly uncomfortable to admit, but it’s real.

When you spend most of your day working with agents that respond instantly, know all the answers, never forget what you said, and work 24/7 — your patience with humans plummets. Fast.

The SaaStr team catches themselves thinking things like: What do you mean you don’t know the answer? My agents know it instantly. Or: What do you mean you forgot? I’ve told you this 10 times already. Or: Why is this taking three weeks when an agent would do it in three hours?

Anyone who has managed more agents than humans notices that it genuinely erodes your ability to deal with people. Agents don’t get impatient when you take a while to respond. They don’t need emotional management. They don’t have career ambitions that compete with the task at hand.

The danger is real: you start wondering if you can just build an agent to deal with a particular person. And sometimes the answer might actually be yes. But the human implications of that shift — for the team, for leadership style, for the ability to collaborate — are significant.

The flip side of that coin: tolerance for lazy work from external vendors has dropped to zero. When agencies send a one-sentence deliverable with a price tag of 40 to 50 thousand dollars per month in 2026 — no detailed plan, no timeline, no AI-powered proposal — they don’t even get a response. If you can’t use basic AI to put together a better pitch than that, what exactly are they paying you for?

There’s no clean answer to this problem. But anyone managing agents at scale needs to be aware that it happens and actively course-correct.

5 Quick Notes Straight From the Trenches

• The 90/10 rule of buy versus build. 90% of your agents should be off-the-shelf third-party tools. 10% built with vibe coding. You only build when nothing on the market addresses your specific use case. SaaStr wished they hadn’t needed to build their own AI VP of Marketing. But nothing else could do what it does.
• Marketo is dying as a source of truth. Legacy marketing automation platforms are atrophying fast. Salesforce now has orders of magnitude more data than Marketo for their operation. The Marketo renewal is coming right after SaaStr AI Annual and will most likely be canceled. Even switching to HubSpot doesn’t make sense at this point — what they need is something natively agentic that pulls directly from Salesforce and creates campaigns autonomously.
• Agent ROI is simple to measure. Attribute closed revenue directly to each agent. Unlike traditional marketing attribution, there’s rarely a multi-touch problem at the agent level. Which agent touched which lead? Did it close? That’s it. The original goal was even simpler: replace a six-figure hire with a five-figure agent and maintain the same revenue.
• Too much data is a real problem. When you have more than 20 agents collecting data, you actually need to limit what they ingest. Knowing that a prospect visited certain pages is useful context. But does the agent need to know that person worked at another company and came to SaaStr five years ago? No. That’s irrelevant context that worsens the output. More data doesn’t mean better — it often means worse.
• Agentic marketing is the big market gap right now. Sales agents are well ahead. There’s still no agent capable of doing full-scale email marketing — segmenting the database, creating campaigns from historical data, and sending to the right contacts at the right time. Not just writing copy. Actually executing. Whoever builds the agentic replacement for HubSpot and Marketo will own a real piece of the next era of B2B.

The Big Takeaways

After facing all the technical and organizational challenges, the picture gets clearer depending on where you are in the journey:

• If you’re putting your first agents into production: know that complexity scales non-linearly. Going from 2 to 20 agents isn’t 10 times harder — but it’s a fundamentally different management challenge.
• If you’re already managing multiple agents: you need daily check-ins with each agent, a planned blackout period for every new addition, and at least two people who understand the full stack. Today. Not next quarter.
• If you’re building B2B products with AI agents: the ROI bar is absurdly high right now. If your agent can generate six-figure meetings in its first week or reliably replace multiple humans, demand will be enormous. Monaco’s demo calendar is booked solid for two months, including weekends. Basis raised funding at a billion-dollar valuation and can’t keep up with demand. If nobody is lining up to buy your agent, your agent isn’t good enough.

At the end of the day, the journey of scaling AI agents in real production is far less about technology and far more about people, processes, and organizational decisions. Unifying contexts, structured onboarding, security applied with discipline, and agent management with clear ownership are the pillars that determine whether an agent operation will generate real value or become yet another project that looked great on paper but didn’t survive contact with reality.

The agents aren’t going away. They’re just getting harder to manage. And that, as strange as it sounds, is the best sign that they’re actually working. 🚀