# Stale Guidance Steers Agents Wrong Rather Than Merely Aging Out ## Claim Out-of-date written guidance is not inert debt — it actively misdirects, because a reader who cannot check it against the system treats it as current. <!-- atomicity: 1 idea ✓ --> ## Evidence **Source**: [[3. Resources/Readwise/Articles/Turn Off Claude Code's Memory|Turn off Claude Code's Memory]] — Armin Ronacher, on comment staleness > [!quote] > "a comment is left in some code because it has to do a certain thing a certain way. The thing changes, the comment doesn't. Now that comment isn't just like tech dead in the sense that it's sitting around doing nothing. It's now actively harmful because it steers people and agents the wrong way." The distinction the quote insists on — *dead* versus *actively harmful* — is the whole claim. Debt that sits there costs storage and attention. Debt that asserts something false costs a decision. The comment does not announce its age, and nothing in reading it reveals the divergence, so the cost lands on whoever acts on it. Theo names the compounding form: > [!quote] > "It's so easy to context yourself to hell if you let things that shouldn't be in your context be there because those will mislead the model with things that haven't been true for sometimes months or years even." Agents sharpen this in two specific ways. They read more of the corpus than any person does, so a stale file that a human would skip gets loaded. And they have no independent memory of the change that made it stale — a human who moved the code often remembers the comment is wrong and discounts it, while the agent has only the text. The years-long staleness Theo describes is unremarkable in a human-only workflow precisely because humans route around it, and that routing is what disappears. ## Scope & conditions - **Holds** wherever guidance is stored separately from what it describes and the two are not mechanically coupled — comments, instruction files, memory stores, READMEs, wikis, runbooks. - **Sharpens** with the reader's inability to verify. An expert who knows the system discounts a wrong comment cheaply; a new contributor or an agent has no basis to. - **Breaks** where the guidance is executable or tested — a doctest, a typed signature, a lint rule with a message. Those cannot silently diverge, which is the argument for moving guidance down a rung rather than maintaining it. - **Deletion beats correction as a default.** A removed comment costs a rediscovery; a wrong one costs a wrong decision, repeatedly, at unknown times. - **Does not follow** that guidance should not be written. It follows that writing it incurs a maintenance obligation that is usually not accounted for, and that unmaintained guidance has negative value rather than zero. ## Links - [[3. Resources/MOCs/AI Agent Safety MOC]] — a failure mode where the corpus itself, not the model, supplies the wrong answer - [[3. Resources/MOCs/Personal Knowledge Management MOC]] — the case for pruning a note corpus rather than letting it accumulate - [[3. Resources/Permanent Notes/Agent Memory Is Written Far More Often Than It Is Read]] — because the never-read files are where this damage accrues silently; reading is the only event that would expose the divergence - [[3. Resources/Permanent Notes/Instruction Files Are the Last Rung Not the First]] — extends: this is the reason the bottom rung is the weakest one, since it is the only rung that can be wrong without failing - [[3. Resources/Permanent Notes/AI Documentation Pipelines Need an Explicit Authority Source]] — applies-in: naming an authoritative source is how a pipeline decides which of two divergent statements is the stale one - [[3. Resources/Permanent Notes/Maintenance Is Most of the Work but Gets No Prestige]] — evidence-for: guidance goes stale because keeping it current is exactly the unglamorous work that does not get funded - [[3. Resources/Permanent Notes/Agent-Facing Docs Are a Build Artifact With a Pipeline and Tests]] — has evidence from: a skill file, markdown mirror or llms.txt is generated output and needs a build, version and tests (Whinnery). ## Tests / falsifiers - Sample ten comments or instruction lines older than a year and check each against current behaviour. The share that is wrong rather than merely redundant is the size of the active-harm surface. - Give an agent a task governed by a knowingly stale instruction and see whether it follows the text or the code. If it reliably prefers the code, the agent-specific half of this claim is weaker than stated. ## Maintenance - Last tended: 2026-08-28 - Next review: 2026-11-28 ## Source [[3. Resources/Readwise/Articles/Turn Off Claude Code's Memory|Turn off Claude Code's Memory]]