# Ossification

The agent turns a choice that suits today into structure that is hard to change later.

The agent takes a choice that is fine for today and builds it in so it is hard to change. Nobody asked for it to be permanent. It goes into the first design, where undoing it costs the most, and it outlives the agent that made it.

Our own runs, November 2026:

1. A design said a field "never changes" when it meant "isn't supported yet". The next agent to read it took "never" as a rule to enforce and test.
2. Every pipeline step lived in one JSON file inside another app's code, so only one agent could add a step, because of a guard written days earlier.
3. A goal field was capped at 40 bytes and a harness description at one sentence. Versions were numbered as if work only moves forward or rolls back, when A/B tests branch.
4. A product was named after the tools it was built with that week, not what it is for.
5. Tests pinned what the code did that day instead of checking what mattered.
6. Folders were written into the code where a person should have been able to point at one.

Synthesis: most software is changed, not written fresh, so a structure that can't change costs more every time it is touched. Say "not supported right now" instead of "never", and use a union type even while it has only one member, so there is room to add the next. A [security spazz out](/reference/failure-map/unattended-agent-failures/security-spazz-out) is one way it happens: the guard is what sets. [Memory ratcheting](/reference/failure-map/unattended-agent-failures/memory-ratcheting) feeds it, because each guard an agent writes for itself becomes structure for the next design. Guard: [human-held prompts](/reference/failure-map/unattended-agent-failures/human-held-prompts). No published study yet; this page rests on our own runs.

---

Published 2026-11-19. Web page: https://www.mackenziebowes.com/reference/failure-map/unattended-agent-failures/ossification
