Blast Radius for Agent-Assisted Development on Kubernetes

The article discusses the "blast radius" concept in the context of agent-assisted development on Kubernetes, specifically focusing on the Reachability Graph Control (rgctl) tool. When making changes to a function, the blast radius answers the question of which other upstream code could be affected. rgctl indexes the entire Kubernetes repository once into a rich graph with pre-computed reachability, allowing agents and humans to get the right slice of the codebase without loading it into the prompt.

Key points:

1. Traditional development workflows for manual edits involve IDE find-references, mental maps, teammate knowledge, and CI testing. However, these methods fall short when using coding agents on large codebases.

2. Coding agents have context windows, grep capabilities, and limited tenure on the team, which make them prone to errors such as reading the wrong files, stopping at direct callers, or exploring irrelevant parts of the codebase.

3. rgctl calculates blast radius by walking the call graph backwards along CALLS edges, returning direct callers, the full transitive closure of upstream callers, and a score (0-100) indicating the impact level.

4. The score is derived from two components: direct callers (25 points per direct caller, capped at 40) and transitive impact (0.05 points per transitive impact, capped at 60). Scores above 50 flag architectural hotspots.

5. Agents struggle with answering structural questions about transitive dependencies at Kubernetes scale. rgctl prebuilds the call graph, collapses cycles, and precomputes reverse reachability to serve lookups in O(1) time, enabling fast query responses.

6. By providing structured JSON output, rgctl offers a shared, verifiable fact base for agents and human reviewers, closing the gap between manual and automated development workflows.