Agentic Slop
If AI agents are so powerful, why aren't we using them more?
If AI agents are so powerful, why aren’t they everywhere yet? We’ve been hearing about “agentic AI” for over a year now. Every startup promises it. Every incumbent is pivoting to it. LinkedIn is full of thought leaders explaining how agents will replace entire teams with digital workers called ‘Bob’ and ‘Mary’.
The demos are compelling. AI agents can clearly do incredible things, so why aren’t we seeing more of them?
My take is:
LLMs and agentic frameworks like LangChain make it relatively easy to build AI agents
But, building agents that behave predictably, reliably, and cost efficiently is a lot harder than most people imagine
The best agents are specialized to a narrowly defined task (e.g. Cursor in coding), but building specialized agents takes a huge amount of work
Under pressure to deliver, teams choose to ship generic agents that sound impressive because they can handle a lot of different tasks. When deployed, these agents fail to deliver at anything close to the level of reliability required for complex enterprise use cases (e.g. Microsoft Copilot)
Seeing these challenges, many teams turned to ‘forward deployed engineers’ (aka implementation consultants) to customize agents for each customer to cover the gaps
These dynamics led to a wave of new AI agents in 2025 that just didn’t-really-work. To get AI agents working, teams should focus on making agents for narrowly defined tasks first, before expanding to more use cases.
It’s clear why AI agents are so promising. An agent can reason through complex, multi-step decisions. It understands context, calls external tools, gathers information, makes judgment calls. It doesn’t follow pre-defined rules and instead thinks through problems like a senior employee would.
The problem is that LLMs are unpredictable. When you ask Claude or ChatGPT a question, you don’t always get the same answer. Now think about what happens when you chain thirty or forty of these calls to LLMs one after another. Now add a dozen different tool calls. The number of possible branches in the decision tree explodes as each step introduces variance and every tool call introduces potential failures. The thing that makes agents so powerful is also their liability. More steps means more that can go wrong. More autonomy means more ways to wander off course.
This is where Agentic Slop comes from. Anyone can take an off-the-shelf agent framework, point it at a vague goal like “help security teams” or “automate IT workflows”, add a dozen integrations, and ship it with a $100k price tag and a forward-deployed engineer to make it work.
The fundamental mistake is trying to make agents do too many things. A general-purpose agent that can “help with anything” has been optimized for nothing. It’s like hiring a consultant who claims to be an expert in “business.”
Building a great AI agent is hard because you have to solve three very specific problems:
1. You need to teach the agent how to plan and think through YOUR specific task.
LLMs are general reasoning engines, but they’re not automatically good at domain-specific problems. You need to encode deep expertise about how a human expert would approach the problem. What information do they gather first? What questions do they ask? What shortcuts do experts know that aren’t documented anywhere? This takes months of iteration with actual domain experts.
2. You need to give the agent the right context and tools.
An agent is only as good as the information it can access. What data does it need to fully complete its task? How do you get that data reliably? How do you format it so the model can actually use it?
Get this wrong and agents confidently give you wrong answers because they’re missing crucial information. Get it right and you need deep integration with your specific systems. Not just the API connections, but thoughtful engineering of what context matters and how to surface it.
3. You need to optimize the cost profile for your specific use case.
Running a powerful LLM model 40+ times per task gets expensive, fast. But using a cheaper model might mean the agent makes more mistakes. The right tradeoff depends entirely on the value of the task and the cost of errors. A general-purpose agent framework can’t optimize this for you.
This might be controversial given it’s the sexiest job title of 2025, but in my opinion there is one clear tell that a company is selling Agentic Slop: they need forward-deployed engineers to make it work.
“Forward-deployed” is a fancy way of saying “our product doesn’t actually work out of the box, so we’re sending expensive humans to do custom engineering at your site.”
I get why companies do this. Customization is legitimately hard and every customer’s environment is different. Palantir has shown that the services+software model can create a big business.
But a truly great specialized AI agent shouldn’t need this. A company that has done the hard work of encoding deep human expertise, engineering the right context, and optimizing for a specific use case should be able to deliver something that works on day one.
AI agents are incredible. The technology is genuinely transformative which is why it’s so frustrating to see such slow progress in 2025. With the exception of a few industries (e.g. coding, customer support) we’re in the slop phase of AI agents right now. Half the market is trying to ship general-purpose frameworks and pretending that professional services make up for product gaps.
The companies that win will be the ones who specialize ruthlessly. Pick one narrow but valuable task and become undeniably great at it before expanding. Invest in deep domain expertise. Work with practitioners to encode how experts think. Nail the context engineering problem. Figure out what data the agent actually needs and build robust pipelines to get it.
Ship products that just work out of the box. If you need consultants to make your agent useful, keep iterating until you don’t. The agents that work will be narrow, specialized, and deeply optimized for their specific task.
Everything else is Agentic Slop.