Cisco automates code reviews with multi-model agent strategies

Cisco engineers are using AI agents to automate pull request reviews, with more than 95% of new AI features at Cisco shipped using Codex, according to an OpenAI case study. The company acts as an early design partner for Codex, having tested previews in October and November of the previous year. Engineers use 'agents.md' files to customize review bots with specific instructions on security, code clarity, and modularity. To mitigate model bias, Cisco employs a multi-model strategy, using different model families for coding and review tasks, including open-source models like Mistral and post-trained proprietary models.

The company also built Cisco Cloud Control, a unified platform integrating over 130 SKUs across 74 products, which was developed in under eight months using agentic coding to accelerate throughput. Sampath advises developers to refactor skill files with each new model generation rather than discarding them, as new models often bake in capabilities that previously required custom context.