Engineering AI agents for on-call and incidents
Engineering AI agents for on-call and incidents
AI has delighted us most when any plausible answer is a good one. Generating a website, drafting copy, sketching a layout: many outputs work, and the model picks one. Production is different. An incident has one right answer, the actual root cause, and a confident wrong one costs revenue.
This whitepaper explores the core requirements of the AI harness needed to build the AI SRE that works beyond an individual user or team.
Model orchestration: route each step to the best model, and absorb every new model release
Context engineering: retrieve the precise context an investigation needs, and nothing more
Causal reasoning: pursue several hypotheses at once and verify each against the evidence
Governed actions: let the agent act on production behind approvals and a full trace
Continual learning: turn every investigation and correction into something the next one uses
Domain evals: measure quality against production-like cases on every evolution of models or agent architecture
