AgenticOps

Client / Project

AgenticOps is an internal proof-of-work for Meta.Dev that validates an agentic on-call layer for small-to-mid engineering teams. We built it for a real client's production environment after they complained that on-call was burning down their senior engineers' attention without proportional reliability gains.

Problem

The team's on-call rotation paged a senior every time something tripped a threshold. Most of those pages were not real incidents — they were noise, transient blips, or known-good behavior that needed acknowledging and dismissing. The cost wasn't the time spent fixing things; it was the constant context-switching cost of triaging false positives.

Approach

Two engineers, six weeks. The agent sits between the alerting system (PagerDuty) and the human on-call. When a page fires, the agent does three things in parallel: pulls the matching runbook (if one exists), checks recent deploys and related metrics for context, and drafts a triage summary. If the page is high-confidence noise (transient threshold breach, known-recovery pattern), the agent acknowledges and suppresses with a logged reason. If it's ambiguous or new, the agent pages the human with all the context already assembled.

We did not let the agent take remediation actions. The contract was: triage and context, not control. Every decision the agent makes is reviewable, and the runbook draft it produces is the artifact the on-call engineer actually uses to resolve the incident.

Outcome

Mean-time-to-acknowledge dropped from 18 minutes to under 3 — most of that gain came from the agent acknowledging known-noise pages without waking anyone. The on-call engineer who does get paged now has a triage summary and the relevant runbook section already pulled up, which compressed mean-time-to-mitigate by about 40%. The team's qualitative feedback was the more important signal: senior engineers stopped dreading the rotation.