autonomous DevOps agents at scale
Linktree's CI optimization agent merges 80+ PRs, cuts build times 60% GDG MelbourneTL;DW
- CI optimization agent merged 80+ PRs across 20+ services; fastest win reduced one pipeline from 22 to 8 minutes (64% improvement).
- Used specialized agents for different tasks: Haiku for pre-screening and review, Sonnet for implementation—matching model capability to task scope reduced costs and improved accuracy.
- Playbooks are markdown files defining trigger conditions and fix recipes; kept small, targeted (e.g., parallelize tests, cache dependencies, retry flaky steps) to enable focused PRs.
- Three-agent funnel: pre-screener identifies violations, implementation agent builds fixes, reviewer agent validates—reduces 120 scanned repos to ~40 merged PRs per cycle.
- Verification loop automatically fixes CI failures by parsing logs, then validates success by checking build times and confirming no regressions in latest builds.
- Created custom MCP wrappers (backend-for-frontend pattern) instead of generic APIs to reduce token overhead calling Buildkite—critical for agent efficiency.
- Auto-research technique: log every agent turn, run another AI to find repeated failures and inefficiencies; discovered agent was unintentionally using sub-agents, fixed after analysis.
- AI cost to AWS bill comparison: optimizing CI runtime reduces EC2 spend far more than agent calls cost, creating net savings plus higher developer productivity.
- Prefer small, targeted PRs with human review over massive changes; all agent work goes through GitHub PR review to maintain transparency and bounded blast radius.
- Tried one large agent doing everything initially—unbounded scope caused token explosion and poor results; specialization fixed both quality and efficiency.
TL;DW
- CI optimization agent merged 80+ PRs across 20+ services; fastest win reduced one pipeline from 22 to 8 minutes (64% improvement).
- Used specialized agents for different tasks: Haiku for pre-screening and review, Sonnet for implementation—matching model capability to task scope reduced costs and improved accuracy.
- Playbooks are markdown files defining trigger conditions and fix recipes; kept small, targeted (e.g., parallelize tests, cache dependencies, retry flaky steps) to enable focused PRs.
- Three-agent funnel: pre-screener identifies violations, implementation agent builds fixes, reviewer agent validates—reduces 120 scanned repos to ~40 merged PRs per cycle.
- Verification loop automatically fixes CI failures by parsing logs, then validates success by checking build times and confirming no regressions in latest builds.
- Created custom MCP wrappers (backend-for-frontend pattern) instead of generic APIs to reduce token overhead calling Buildkite—critical for agent efficiency.
- Auto-research technique: log every agent turn, run another AI to find repeated failures and inefficiencies; discovered agent was unintentionally using sub-agents, fixed after analysis.
- AI cost to AWS bill comparison: optimizing CI runtime reduces EC2 spend far more than agent calls cost, creating net savings plus higher developer productivity.
- Prefer small, targeted PRs with human review over massive changes; all agent work goes through GitHub PR review to maintain transparency and bounded blast radius.
- Tried one large agent doing everything initially—unbounded scope caused token explosion and poor results; specialization fixed both quality and efficiency.
A multi-agent system autonomously identifies pipeline bottlenecks and submits fixes via GitHub PRs, achieving a 50% acceptance rate across 20+ services. Specialized agents handle pre-screening, implementation, and review; switching from a generalist to specialized architecture cut token costs, and one pipeline dropped from 22 to 8 minutes.