# LinkedIn Post: The U-curve of AI agents

> Source: docs/research-fable-agentes-curva-u.md — disclosure version, 2026-08-02

---

Your AI agent doesn't get lost because the repo is large.
It gets lost because the repo isn't yet large in the right way.

We've all seen the pattern:

📈 Small project: the agent flies. Every task ships on the first try.
📉 Repo grows: at some indeterminate moment, the agent starts wandering, touches files it shouldn't, and one day proposes a complete rewrite.
📈 Software matures — isolated modules, extensive tests — and the agent regains its bearings. With an even larger repo.

That's not a "more code = worse" curve. It's a U-curve. And a U-curve clears the repo size as the culprit.

The real variable is different: how much *verified* context does the agent need to touch something without breaking anything. In the small repo, it has everything. In the mid-sized repo without structure, coupling means to touch A you must understand B, C, and D... and everything the agent didn't read it fills in with assumptions that *sound* like knowledge. The erratic paths are exactly that: sampled guesswork where evidence runs dry.

And why does it want to rewrite everything? Because it's its way to reduce uncertainty: its own code is the only 100% consistent thing inside its window. Locally rational. Globally destructive.

And why do tests rescue it? Because a failing test is not another instruction competing for attention: it's evidence that collapses the assumption. Tests turn errors that accumulate into errors that get corrected.

The uncomfortable conclusion for 2026:

Isolated modules don't make the agent smarter.
They make its ignorance irrelevant. Which is better.

The best AI-readiness strategy didn't come out this year. It came out in 1972, it's called information hiding, and David Parnas wrote it. Modularity, clear boundaries, dense tests. What was always good engineering is now, also, the infrastructure your agents need to stay on course.

And the valley — that in-between phase where structure doesn't yet exist — is where most agent-driven projects die today. There, feedback has to be imposed from outside: contracts, deterministic verification, tests before you "have time for tests."

Where on the U-curve is your repo?

#AIAgents #SoftwareEngineering #DeveloperExperience #CodeQuality

---

## Usage notes

- Alternative hook (more confrontational): "The best AI-readiness strategy is 54 years old and no AI lab invented it."
- Length: ~1,900 characters, within LinkedIn optimal range (breaks at "see more" after ~210 characters: the two-line hook stays intact).
- For citing evidence: 2026 benchmarks report ~91% → ~15% pass-rate collapse as repos grow without structure (see research-fable-agentes-curva-u.md §C1, with provenance caveat).
- English version available on request.