Agents for infrastructure operations

Incident triage in Discord and change approval with AI risk assessment.

My role

I wrote the triage bot from scratch (9 of 9 commits) and the approval gate.

7
change categories: 3 min
cap per investigation

SPAN 01 // CONTEXT

Problem

Incidents landed in Discord and someone had to search Jira, copy logs and open a ticket. Infrastructure changes had no auditable approval step.

SPAN 02 // IMPLEMENTATION

Solution

A Go bot that reads the thread, investigates with Gemini, queries Jira over MCP, avoids duplicates and opens or updates the ticket. Plus a change gate where a request, made over MCP from the editor, gets a risk review through Bifrost, becomes a Jira RFC and shows up as an approval card in Discord.

SPAN 03 // DETAILS

Highlights

A custom tool-calling loop over stdio MCP, because LangChainGo's stock agent was dropping arguments.
Per-thread locks, MCP reconnects and a 3-minute timeout with recover.
The gate blocks self-approval, requires a reason on rejection and attaches the full diff.
Before the RFC is opened, the AI checks citizen impact, data and IAM exposure and whether rollback is feasible.

SPAN 04 // OUTCOME [200 OK]

Result

The triage bot runs in production on IplanRio's Kubernetes.