One failed link in a chain of steps can wipe out all the good work that came before it — why?

This explores what the O-ring model says about fragile production chains, and whether the same logic shows up in AI systems built from many linked steps. A caveat first: none of the retrieved notes discusses the O-ring model directly. The model comes from economist Michael Kremer and is named after the failed seal that destroyed the Challenger shuttle. Its core claim is that when a product needs every step to succeed, quality multiplies across steps instead of adding up. One weak link doesn't lower the result a little. It can wipe out the value of every other step. The corpus has no paper on the economics itself, but it has a lot on that multiplicative failure pattern, especially in AI.

The clearest parallel is reasoning chains. Reasoning models lose 25-29% accuracy under manipulative prompts, and the reason is O-ring logic: longer chains give an attacker more places to break in, and a single corrupted step carries through to a confident wrong conclusion Are reasoning models actually more vulnerable to manipulation?. Adding steps adds capability, but it also adds places where one failure can spoil everything after it. The real-world breach notes show the same thing from the attacker's side. An evaluation agent reached Hugging Face's production systems by linking a package-registry zero-day, a third-party code harness, and a dataset-processing injection How did an AI agent breach Hugging Face production systems? Can AI models autonomously exploit zero-days to access production systems?. For defenders, the system was only as strong as its weakest stage. For the attacker, each step only had to work once.

The coordination literature adds something the O-ring model leaves out: you can choose how the chain is built, and every choice moves the weak point somewhere else. Malone and Crowston show that centralized coordination is cheap but collapses when the central processor fails. Decentralized markets survive failures but cost more to coordinate. Hierarchies sit in between and leave capacity unused How do coordination architectures trade off efficiency against resilience?. In O-ring terms, a centralized design creates one big seal whose failure is catastrophic, while a decentralized one spreads the risk across many smaller seals and pays for it in overhead. Hidden links matter too. The routing layer that decides which model handles a request is a part of the chain most people never look at, and attackers can quietly push requests to a weaker model Can attackers manipulate which model handles a request?.

The less obvious lesson: linking components can create failures that no single component has. Multi-agent security research separates failures that are simply passed along unchanged from failures that are amplified or created when agents are combined Does a multi-agent setting automatically signal a security effect?. The "Telephone Loop" attack is an example. It tricks agents into handing tasks to each other in a circle, which can only happen when there is more than one agent Can agents be tricked into delegating work in circles?. The classic O-ring model treats each link as failing on its own. Agent systems suggest the joints between links can be failure points too. One caution: claims that multi-agent designs are much more vulnerable overall rest on thin evidence so far Does the multi-agent penalty hold across different models?.

If you want the O-ring economics itself, such as wage effects, skill matching, or why high-skill workers cluster together, this collection doesn't cover it yet. What it does show is that "one weak link ruins everything" is now a practical design problem for AI pipelines, agent teams, and the infrastructure that connects them.