"Blast radius" is one of those operational concepts that everyone nods along to and almost nobody actually designs for. It's the answer to a simple question: when this thing fails — or when this credential leaks, or when this agent goes rogue — what else goes down with it? The term comes from SRE culture, but it's having a second life in the era of AI agents, IaC pipelines, and org-wide service accounts that quietly hold god-mode across half your infrastructure.

Based on the title and the personal-blog format, this post likely walks through blast radius as a design lens: how to shrink it, how to measure it, and where engineers reliably underestimate it. Expect concrete territory like:

• Credential scoping — the difference between a token that can read one bucket versus a token that inherited the whole org's IAM tree because nobody wanted to write another policy.
• Deployment topology — single-region vs. cell-based architectures, and why "we'll just fail over" is a plan, not a design.
• Human blast radius — the on-call engineer who can kubectl delete anything in prod at 3am because MFA-gated break-glass was "too annoying."
• Agent blast radius — the newest and scariest flavor: an LLM with tool access whose effective permissions are the union of everything it can call, chained.

Why it matters for a technical audience right now: the industry is enthusiastically handing production keys to autonomous agents while still using the same coarse-grained IAM patterns that made the 2010s a parade of "one leaked AWS key deleted the company" postmortems. Every new MCP server, every new agent-with-tools integration, is a blast-radius decision — and most teams are making it implicitly by copying an example config.

The really useful version of this post is the one that gives you vocabulary to push back in design review: "what's the blast radius of this service account?" is a question that reframes an argument about convenience into an argument about consequences. That's a rare gift in an industry that mostly measures "did it ship."

Zero comments and one point is criminally low for something that touches SRE, security, and agent design simultaneously — three audiences that would all benefit from reading the same essay.