## 0003 / 2026-08-28

Why aren't there more agents?

You could have 1,000 agents running at all times, doing useful tasks for you all day long. However, you almost certainly don't have this many agents working for you. Why?

The models are good enough. The harnesses are good enough. There is plenty of APIs everywhere, so... what's missing?

First, let's talk about the Human limits.

We simply cannot keep 50+ different agent sessions in mind, at all times, and unblock all of them by hand. No matter how you try, you will reach your limit soon.

You forget which agent is doing what. You lose track of the ones waiting on you. And every agent eventually stops and needs you -- a permission, an answer, a decision -- so you become the bottleneck. These compound: you forget the waiting ones, so they sit idle for hours.

It's like having 50+ direct reports in a company. There's a reason why companies have 6-12 employees per manager.

One of the most productive engineers at Cursor (by output, by PRs) talked about how being a good manager translates to being good at agents. So this is not just our imagination. This is validated by people who are more locked-in than us.

The second reason: people don't know where to build an agent. They either don't know how to set it up -- Grok Bot was a huge step in the right direction, making it easier for everybody -- or they don't know what is a good task for an agent.

There have been studies on this, on the human brain -- what percentage of things we do throughout the day are forms of habit, and what are totally new. And it's like 95 to 5.

95% of the things we do throughout the day are things we've done before, things we already know how to do. We're not consciously, deeply thinking about them. Usually only 5% is the kind where you have to figure it out from scratch.

For example, going to a new location. Where is this building? How do I get here? That's why these things take a lot of mental effort, because it's a lot of figuring out. Whereas most tasks are routine. Usually you do the same thing, or you do it only slightly different.

Over 90% of the tasks we, humans, do on a daily basis are inherently repetitive and delegatable to agents. The units of work are everywhere. People just haven't converted their workflows, their SOPs, into agent skills.

What is a good unit of work for an agent?

That is a priceless skill to have. Knowing where to put an agent TODAY, vs what is not possible today, maybe a year from now, vs what is not even worth automating, is an invaluable ability to have.

If your idea to "i'm gonna put an agent here" it's too ambitious, the agent probably cannot do it well, and then people have a bad experience with AI. Identifying the right unit of work for an agent is huge.

The third reason: context. Obviously you need to use the right models -- frontier models or near-frontier models, not some dumbass 7B model. But if you have the right model and a solid harness, what's limiting the agent is context and connectors. Mainly the context. Connectors are only a first-time issue. Context is an ever-present issue.

A lot of people don't even describe how to do a task. They don't give the agent enough info. And then they're pissed that it didn't do it exactly like they wanted to.

Some of the tactics we use at the David Ondrej team include: transcribing calls, making sure Fireflies is on every meeting, doing Looms, forcing yourself to teach it. Shared repos, like our meta-learnings repo. These are good tactics for solving the context issue. For connectors, there's DeepAPI and Composio.

The fourth reason: a lot of agents are just not aggressive enough. A lot of people set up their agents as "if there's something unexpected, message me." This is a loser strategy. This is fundamentally betting against the model.

Grok Bot is super proactive, which is good for getting shit done. Sometimes it can be annoying, but on average I would say it's good, because if we don't set this up, no work is happening. The Cursor developers wrote the system prompts well, I think.

Maybe the problem is dynamic... if the user didn't give you enough to do the task, be very proactive. Because once I have bots that have a clear task, they're not as annoying. That's the biggest risk: the user doesn't do the setup, and the agent never gets the opportunity to even do any work.

At the same time, you don't want the agent doing shit you're not aware of. So how do you make it so you trust the agent more? You really need to have a clear idea of what the agent does and what it doesn't do. A clear reason why you created the agent. It shouldn't be "oh, I think I need another agent in my life or business." That's complete bullshit. That's the mistake of hiring an employee without knowing what role you're hiring for.

That's how people should think about it: you're hiring this agent. There's a cost to it, because he's consuming your plan. So what are you hiring him for? What do you want him to do? If you're hiring for accounting automation, you probably don't want it brainstorming marketing ideas. Don't make the agent too general and too broad.

One of the reasons Grok bots are succeeding is they're specialized. People have different Grok bots running for different things, and the user interface makes it very clear that this is not a single agent. That's the issue with Hermes and OpenClaw: it's a single agent, and you have to kind of mix the context. With Grok bot, it's very clear on the UI that you should have multiple agents. The interface makes it obvious, and it's very easy to create a new bot.

The fifth reason: persistent environments. For agents to exist, they cannot be dying. It's like, why aren't there more humans? Humans are dying. It's the same thing.

So what makes an agent die? Its environment times out. There's some random error. Maybe the task doesn't become relevant. Or there's an error and it doesn't self-heal. This was a massive issue in the early days of OpenClaw... you'd deploy it on a VPS, and then a few days (or weeks) later it would just stop working. Or, the agent began failing on a task that it used to do 100+ times without problem, due to some random update.

This is a big issue. Why aren't there more agents? They're not reliable enough. Not consistent enough. Not predictable enough. Part of that is the issue of the model, but I think the bigger part is the environment, the harness, everything around it.

The models are good enough. The harnesses are good enough. Everything around them isn't.

All articles