Root Minds · Cisco Innovation Hackathon, top 10 of ~100

The problem

When something breaks, engineers normally jump across logs, metrics, alerts, and dashboards, piecing the story together by hand. Root Minds runs the investigation for them. It gathers the relevant signals, runs approved checks on the device to confirm what it finds, and points to the root cause.

The guardrail

• Specialized agents. Went with a multi-agent design instead of one model trying to do everything, so each investigation stage, log reading, metric correlation, code tracing, final summary, is a separate agent tuned to that one job instead of a single prompt trying to cover all four.
• Selective LLM calls. At enterprise scale, an LLM call at every step gets expensive fast. Routine log and metric patterns resolve deterministically with pattern and regex matching before the model is ever called, so LLM calls are spent only where real judgment is needed, like cross-source correlation and code-level root-cause tracing.
• Approval gate. Root Minds runs on a shared staging environment used by many testers, so it can't fire commands at a device on its own. When it needs more evidence to confirm a finding, like extra logs, metrics, or a show command, that action is allowlisted and gated behind an approval. Nothing touches a device until someone approves it, which keeps an autonomous agent safe in a shared, production-like setup.

Stack

Multi-agent AI · LLM orchestration · Prometheus · Loki · Grafana · Kubernetes