Lessons from the Telegraph for LLM Agent Design

This started as a question applied historians ask routinely: can we find lessons from history that inform current practice? The specific case was commercial telegraphy and LLM agents — two systems that compress language for expensive, noisy channels.

I went through a lot of telegraph techniques and operations asking what might be useful when working with LLMs. Almost everything survived — delimiters, confirmation, priority tiers, compression, validation, observability, access control — but as information management fundamentals, not telegraph-specific insights. They entered the general toolkit long ago. (Some comparisons are in the appendix.)

But there’s one area that might not have! A lessons from history? I give you: redundancy in control labels.

What codebooks got right

Telegraph codebooks had to compress messages cheaply while ensuring a damaged codeword didn’t silently become a different valid one. Bentley’s rule: good codewords should differ by at least two letters.¹ The ABC Code applied the same principle, screening out Morse-similar words.² Bellovin surveys the full design space — compression, error correction, confidentiality — and how codebook designers balanced all three.³

Good compression leaves enough redundancy to detect mistakes.

This didn’t carry forward into LLM agent design because the problem is newly common. LLM agents combine properties that make label confusability newly dangerous:

• probabilistic language generation (the model can produce near-miss strings)
• exact symbolic control (downstream code treats labels as exact inputs)
• natural-language prompts wrapped around short labels (easy to miscopy or misread)
• tool calls and routes with real side effects (wrong label → wrong action)

Recent research confirms this is not hypothetical. Qiu et al. show that LLM tool selection is sensitive to surface-level cues in tool names: assertive names, specific phrasing, and similar-looking alternatives all shift routing behavior measurably.⁴ Wang et al. demonstrate that small adversarial perturbations to tool names can reliably redirect tool calls to attacker-controlled tools.⁵

Example labels that are too close:

A1  /  AI  /  Al
prod  /  production  /  live
docs.api  /  doc.api
FACTCHECK_API  /  FACT_CHECK_API

In an agent system, these may route work, call tools, select environments, or satisfy exact enum values. A wrong valid label is worse than an invalid label, because invalid labels fail validation, wrong valid labels pass validation and trigger the wrong action.

An agent-aware check

The check that matters is not just string similarity. It is string similarity weighted by what the label controls.

Generic similarity check:

docs.api / doc.api — edit distance 1.

Agent-aware check:

CRITICAL  docs.api / doc.api
route-name collision across different effects.
read-only route is one edit away from external-write route.

The severity boundary depends on the label’s role:

• Two confusable read-only routes: annoying but low risk.
• Two confusable routes where one is destructive: critical.
• Two confusable environment labels that both mean production: critical.
• Two confusable enum values the model must emit exactly: high risk.

I am experimenting with small tools to check this: wireword. The rules are simple and include visual confusables, edit-distance-one pairs, case-only and punctuation-only differences, plural/stem collisions, and production-alias overload.

Appendix: telegraph practices and LLM-agent counterparts

Sources