New Framework Lets Services Diagnose Their Own Failures at Runtime

An engineer built the Cognitive Autonomic Framework (CAF) after noticing that autoscalers only react to CPU or memory pressure and have no model of what an application actually does. CAF's core idea: at build time, a compiler walks the AST to extract a small Runtime Semantic Topology (RST) graph, capturing service dependencies, failure domains, retry policies, and a closed list of permitted repairs. This graph ships inside the container, so when something breaks the system reads its own architecture instead of reconstructing it from logs.

The design runs on three tiers: millisecond-scale reflex circuit breakers, cheap node-local agents that consult the self-model to attempt bounded fixes, and an expensive central LLM invoked only for rare, hard cases off the critical path. Nodes gossip beliefs rather than heartbeats — if neighboring services independently suspect the same fault, confidence rises, letting the system distinguish a systemic issue from an isolated one without central coordination. Safety comes from structure: agents can only choose from a pre-declared repair list, and any proposed fix must pass verification in a shadow environment before reaching production.

In a working prototype with six real services, giving the reasoner its self-model roughly doubled root-cause detection accuracy, and the improvement came from the structure rather than model size — gpt-4o-mini matched gpt-4o. The safety boundary rejected every disallowed repair. The author is upfront that fleet-scale and long-term validation are still pending.

The takeaway for engineers: architectural knowledge that already exists at build time doesn't need to be discarded and re-guessed by expensive LLMs parsing logs — carrying it forward with the service can make failure diagnosis both cheaper and more reliable.