More and more developers report success with agentic coding.

They use SPARC, Cursor Rules, Claude Code Skills, or simple prompting. The approaches are completely different, yet they all “work” in the sense that they provide code that compiles and that meets specification.

We all see that the models are consistently getting better. Better at instruction following, tool use, context windows, reasoning. The floor has risen.

But we can still fail, and we often do. By failure we meam anything from the failure to deliver code that addresses the intent of the developer to failure to deliver code that is robust and maintainable over time.

Given the power of the AI models, why do some teams still fail? And which approaches will survive the next model switch?

Ingo Eichhorst suggested a crucial connection in his JavaLand 2026 keynote: an LLM is a “noisy channel” in the sense of Shannon’s information theory. The framework proposed by Eichhorst can be used to explain why some approaches work and others don’t, regardless of the model.

When it comes to transmitting information over a noisy channel, Shannon points that there are exactly two levers that allow us to increase throughput:
1. Improve the signal.
2. Improve error corrections. (this means identifying the errors and correcting them)
We can apply the same mode of thinking to LLMs. In order to get better quality answers from an LLM, we must:
1. Improve the “signal” – that is to provide more relevant context so that the LLM can provide a better answer.
2. Identify the errors in the output and correct them.

In practice, we suggest the following:
Improve the signal using Semantic Anchors. These are well known terms that can invoke entire knowledge domains.
Improve the error correction using Redundancy. Have agents, linters, compilers and so forth cross check each other.

Combining these levers form a harness, and experimentation shows: the difference between the best and worst harness is greater than the difference between the best and worst model.

The talk demonstrates how Semantic Anchors, Semantic Contracts, and Eichhorst’s Principle work together in practice, using three fully AI-generated open-source projects.

Takeaways:
– Better models raise the floor, but the ceiling depends on the harness
– Semantic Anchors compress domain knowledge into a few tokens
– Better tests beat better prompts
– Semantic Contracts extend Anchors with project-specific knowledge
– Eichhorst’s Principle provides a framework for investment decisions: harness first