The world of artificial intelligence is moving faster than ever. This is especially true with the rise of agentic AI, which refers to systems that can plan, act, and operate on their own. Companies are eager to adopt these systems. However, as excitement builds, an important question arises: what actually works in production?
Behind the headlines and multi-billion dollar valuations, we are seeing a familiar pattern: hype is outpacing real results.
Hype vs. Reality: The Valuation Disconnect
Consider LangChain. Once a star of the agentic AI scene, a recent report valued the company at over $1 billion, even though its estimated annual recurring revenue (ARR) is below $15 million. That’s a 66x multiple.
- Is LangChain already dead?
- Why is LangChain so hard to use?
- Why are developers leaving LangChain?
The data confirms these worries:
- MIT (2025): 95% of AI projects fail to meet expectations.
- Rand Corporation: AI-related projects fail at twice the rate of non-AI projects.
- S&P Global: The rate of abandoned AI projects doubled from 2024 to 2025.
- This does not mean AI is broken. It shows that most systems today are still experimental. The vision exists, but the reliability is missing.
Why Agentic AI Is So Hard
Agentic systems are not just smart scripts or API chains. They are stateful, long-running, concurrent distributed systems with added complexity:
- Agentic systems are not just smart scripts or API chains. They are stateful, long-running, concurrent distributed systems with added complexity: In short, this isn’t microservices 2.0. It is distributed systems with higher stakes and more complicated failure modes.
- Stochastic Components: Large language models (LLMs) bring probabilistic behavior into execution paths, which requires resilience and new testing methods.
- Specialized Infrastructure: Relying on GPUs for inference adds challenges to orchestration on top of CPU-based workflows.
- Long-Lived Workflows: Agents often run for hours, days, or even weeks, requiring resumability and persistence.
- Side-Effecting Actions: Agents do more than analyze data; they also send emails, update records, and make changes that cannot be undone.
Frameworks like LangGraph (part of LangChain) try to address agent orchestration by introducing durable execution engines. However, let’s be clear: durable execution isn’t easy.
