The RatioA weekly newsletter on reliability economics | | |
|
The Number
9.84 vs 0.77
Critical, revenue-generating systems carry a median reliability ratio of 9.84 while internal tools sit at 0.77 — meaning teams are proportionally over-protecting the systems that matter least and under-protecting the ones that cost most when they fail.
Tier 1 systems, the ones facing customers, generating revenue, and meeting regulators, have a median reliability ratio of 9.84 across 53 responses. Tier 2 systems, internal tools and back-office, have a median ratio of 0.77 across 42 responses. Below 1.0 means you spend more on prevention than failures cost you. Above 1.0 means failure costs are outrunning what you spend to prevent them.
Neutral
The prevention budget exists. It's just aimed at the wrong systems. Teams are spending proportionally more to protect internal tools than the services that bring in money and face regulators. That's like a hospital running its strictest sterilization protocols in the break room while the operating theater gets a once-a-day wipe. Nobody decided this. It accumulated. Fixing it doesn't require new headcount or new tooling. It requires one reallocation decision.
| Internal tools get the prevention budget. Revenue-critical systems run failure costs at nearly 10x what's spent to prevent them. |
|
|
This Week in Reliability
AI Agents Need Production Context
The conversation has shifted from whether AI can help with SRE work to what data sources agents need to be reliable. Runtime context, live production state, and grounded verification are becoming table stakes for agent deployment — because probabilistic models without operational facts create new cleanup costs.
|
Lightrun · Primary evidence Lightrun's MCP skill automatically places conditional snapshots on running processes when monitoring tools capture qualifying errors, mapping failing code paths and capturing runtime evidence at exact lines before raising fix PRs. The workflow follows: Model decides → Runtime Context verifies → Action executes. Prevention This is the pattern that makes agent-driven remediation credible in production: probabilistic reasoning gated by deterministic runtime verification. Without the verification step, you're shipping model hallucinations into your deployment pipeline and budgeting cleanup labor you can't predict. | Agents stay probabilistic; runtime context makes them safe. |
|
Anyshift · Quantified evidence Anyshift's live benchmark showed that resolving infrastructure relationships before model inference cut token usage by 83.5% while reducing response time and estimated cost, without changing task outcomes. Unified production context delivered cheaper, faster agent execution. Prevention Context pre-processing isn't philosophical — it's a cost lever. When you resolve relationships in a graph before feeding tokens to the model, you compress the inference workload and shrink the blast radius of incorrect assumptions. The prevention ROI is in the tokens you don't burn and the rollbacks you don't execute. | Context consolidation cut agent cost by 83.5%. |
|
HackerNoon · Vendor framing Agentic SRE combines AI reasoning, continuous learning, and systemic quality practices to shift from reactive incident response to proactive prevention, helping teams stay ahead of complexity rather than drowning in alerts. Neutral Without production context and verification, 'agentic SRE' is just alert automation with a bigger vocabulary — it accelerates diagnosis but doesn't prevent the next page. | Prevention requires context, not just reasoning speed. |
|
Lightrun · Code review angle CodeRabbit and Greptile analyze code that has never run, regardless of codebase context. Lightrun gives AI access to real production traffic samples and simulates PR logic against them as structural differences. Prevention Static analysis of AI-generated code optimizes for merge speed; runtime simulation optimizes for deploy confidence — different cost centers, different budgets. | Runtime simulation stops bad AI code earlier. |
|
Anyshift · Developer tooling The open-source Anyshift agent plugin gives Codex, Cursor, VS Code, and GitHub Copilot read-only access to current production context when reviewing infrastructure changes. Prevention Read-only production access for IDE agents turns them into informed reviewers instead of confident guessers — you gate write actions until context confirms the change is safe. | Production context turns agents into informed reviewers. |
|
Dash0 · Tool landscape Agent0 absorbs analytical work that slows humans down and turns raw telemetry into narrative. Dash0 emphasizes openness, portability, and understanding over black-box automation for AI-assisted reliability engineering. Firefighting Telemetry-to-narrative translation only creates prevention value if the narrative is grounded in real system state — otherwise you're paying for better incident storytelling, not fewer incidents. | Narrative tools don't prevent unless context grounds them. |
|
|
The Crowd Favorite
- Master of Puppet — Metallica — One upstream timeout without circuit breakers, and the whole stack goes down. Cascading dependencies are not a design pattern. They're a trap.
- Fortunate Son — Creedence Clearwater Revival — Five-nines SLOs applied uniformly across every tier. Sounds disciplined. Actually just inflates on-call load without protecting the services that matter.
- Bohemian Rhapsody - Remastered 2011 - Queen Six movements, zero predictable structure, somehow it holds. That's what good architecture looks like from outside. The 33:1 teams think theirs does too. Freddie rehearsed. They're improvising in production.4. Drop It Like It's Hot - Radio Edit - Snoop Dogg, Pharrell Williams — Error budget hits zero — drop the deploy, not the standards. Every ship after burndown is borrowed reliability you repay in pages.
- Sky and Sand - Paul Kalkbrenner — Nine minutes of slow build before the drop. Proper capacity planning. The teams that burn their budget by day five shipped on the feeling, not the signal.
Prevention
| Reduces future firefighting |
|
|
The Challenger — Ask Me Anything
"Our CTO says our 0.8 ratio is 'good enough for our size.' How do I push back?"
Across 99+ services, teams near 10x aren't bigger than yours. They moved money from reactive work into prevention. Their ratio is twelve times better. Headcount doesn't explain it.
Don't argue the number. Price it. Pull the actual cost of your last three major incidents — engineer-hours, customer credits, executive time. Average those. Multiply by frequency. Put that next to your prevention spend.
That comparison turns "good enough" into a budget line your CTO has to defend out loud.
Neutral
| The data to make the case |
|
|
The Ratio is a weekly newsletter by Florian Hoeppner. Early edition, you are one of the first.
Forwarded this? Get The Ratio in your inbox every week.
|
Take the assessment → reliabilityeconomics.com/benchmark Reply to this email with your take.
How this newsletter gets better: Every link is tracked and every vote is counted. Sections that get a down vote get replaced. Topics that get an up vote stay. No personal data is stored — just click and vote counts per section. Your feedback shapes the next edition.
|
|
|