Reliability Economics

Archives
Log in
Subscribe
May 19, 2026

THE RATIO #2: Your AI agents move fast. Your budget doesn't.

The Ratio #2

Issue №2May 19, 2026

The Ratio

A weekly newsletter on reliability economics

The Number

1 of 56 tech orgs near-optimal (1.8%)

Only 1 of 56 Technology respondents is near-optimally invested in reliability, compared to 7 of 32 (21.9%) in Financial Services, a 12× calibration gap between the industry that built SRE and the one that didn't.

Only 1 of 56 Technology respondents (1.8%) is classified as near-optimal in reliability investment. In Financial Services, 7 of 32 respondents (21.9%) are near-optimal. The remaining 55 tech respondents split almost evenly: 27 under-investing, 28 over-investing. Financial Services shows the inverse pattern: 22 under-investing, 7 near-optimal, 3 over-investing.

The Ratio Take

The tech industry's 27/28 split looks less like a managed discipline and more like a boom-bust cycle — starve the reliability budget, get burned by an incident, overcorrect, repeat. Regulation and audit pressure appear to accomplish what SRE culture alone has not: anchor organizations toward the optimal band. Tech built the vocabulary of reliability economics but has not translated it into calibrated spending behavior. Financial Services is 12× more likely than Technology to be near-optimally invested in reliability — despite inventing none of the tools.

Did this land?👍 More like this👎 Switch it up

This Week in Reliability

Machine Speed, Human Budget

AI agents are shipping code, flipping infrastructure flags, and cascading failures faster than detection cadences, attribution tooling, or prevention budgets can track. The spending mismatch is no longer theoretical — it's appearing in post-mortems, architecture critiques, and governance retrofits this week.

Deep Reads

1

The clean-up cost of AI-generated code is what the velocity narrative leaves out

The New Stack  ·  Economic blind spot in the velocity narrative

The article surfaces downstream debugging, rework, and incident hours generated by AI coding tools — costs that don't appear in the throughput or velocity metrics used to justify AI coding tooling budgets.

The Ratio Take

Every velocity dashboard celebrating AI code output is missing a corresponding reactive-spend column; the cleanup arrives in someone's incident queue, not on the shipping chart that justified the tooling purchase. Budgets that anchor on velocity stats alone are systematically under-investing in prevention while the ratio silently worsens with each agent-generated commit. Velocity dashboards measure output; they don't bill for the cleanup.

2

Why agent harnesses fail inside cloud-native systems

The New Stack  ·  Root cause of machine-speed cascade failures

Agent harnesses break down in distributed cloud-native environments because feedback loops are fragmented across services — by the time a harness registers a bad agent action, the cascade is already several hops deep.

The Ratio Take

The distributed feedback problem is the technical explanation for why blast radius expands faster than detection can catch it: harnesses built for single-service boundaries can't contain an agent whose actions propagate across the mesh before any signal returns. Prevention spend that doesn't account for multi-hop feedback latency is buying a guardrail with gaps already in it. No harness catches a cascade that's already three hops out.

Signals

3

Why agentic AI development needs reliability guardrails

Gremlin Blog  ·  Vendor prevention response

Gremlin argues agentic AI compounds reliability risk because autonomous systems can trigger failure paths humans never designed test cases for; the post outlines blast-radius testing, fault injection, and pre-production guardrails as the structural response.

TakeThe economic case for a guardrails program priced against the cost of one uncontrolled agent action looks cheap in any spreadsheet — the harder sell is that this spend appears as a line item before anything breaks, which is exactly when budget committees push back. Test the blast radius before the agent finds it for you.

4

What kubectl debug doesn't tell you: The silent evidence gap

CNCF Blog  ·  Attribution tooling gap — human-speed forensics vs machine-speed failures

A kubectl debug session can contain the only direct observation of a failing system state; Kubernetes discards the session's termination context once it closes, leaving no durable record for post-incident analysis.

TakeAttribution tooling is still running at human debug-session speed — when evidence evaporates before the post-mortem even starts, every hour of root-cause investigation costs more and produces less confidence in the answer. Evidence vanishes before the post-mortem even opens.

5

How to Detect Terraform Drift Across Multi-Cloud

Anyshift.io  ·  Detection cadence gap — direct embodiment of the seed trend

A development RDS instance had its publicly_accessible flag changed on a Friday afternoon; a once-per-weekday drift detection cadence meant 60+ hours elapsed before discovery, and the post walks through an audit-log subscription architecture that would have surfaced the change in under two minutes.

TakeOnce-daily detection cadences were sized for human-made changes — at agent deployment speed, a 60-hour detection window is a liability that belongs on the risk register and the tooling budget before an agent ever inherits those same credentials. Human-speed detection windows don't fit machine-speed infrastructure changes.

6

A decade of governance: Cloud Custodian at 10 and its role in the agentic AI era

CNCF Blog  ·  Prevention tooling being adapted for the agentic era — 6 days old at publication, within 7-day window

Cloud Custodian, a 10-year-old CNCF-incubating open source policy engine, is being positioned as a runtime constraint layer for autonomous agent actions across AWS, Kubernetes, and infrastructure-as-code pipelines.

TakeRetrofitting a decade-old policy engine to constrain autonomous agents is the prevention budget line item most teams haven't written yet — the gap between when the agent shipped and when the guardrail caught up is exactly where the blast radius lives. Ten-year-old prevention tooling meets 2026 machine-speed agentic risk.

7

How to Trace a Production Incident Back to the Commit

Anyshift.io  ·  Attribution cost multiplied by machine-speed multi-repo deployments

The post describes a 25-minute attribution delay on a Friday-morning page caused by a responsible commit sitting in another team's repo, and presents a four-command sequence for cross-repo blame traversal when git log returns nothing on the first service checked.

TakeWhen AI agents ship code across dozens of repositories simultaneously, the 25-minute manual attribution cost this post documents becomes a recurring line item that compounds with every agent deployment cycle — reactive spend that grows in direct proportion to agent velocity. AI agents commit across repos; your blame tooling still thinks in one.

Did these land?👍 More like this👎 Switch it up

The Crowd Favorite

  1. Everlong — Foo Fighters. Four minutes, the drums never miss. That's what 99.99% feels like underneath everything else.
  2. Under Pressure - Remastered 2011 — Queen, David Bowie. Two vocalists trading the lead. Graceful failover. Load shifts before either voice cracks.
  3. Mr. Blue Sky — Electric Light Orchestra. Redundant harmonies stacked on top of each other. Pull any one track, the song still ships. Defense in depth.
  4. Karma Police — Radiohead. The outro loops until it degrades into noise. Retry storms do exactly this without exponential backoff and a circuit breaker.
  5. Tom Sawyer — Rush. A 7/4 bridge inside a 4/4 song. Most outages live in the odd-time bar nobody rehearsed.
Blast radius is a design choice
The playlist — keep or kill?👍 Keep the crown👎 Dethrone it

The Challenger — Book Review

Seeking SRE, edited by David N. Blank-Edelman (O'Reilly, 2018).

Coburn Watson’s “Context Versus Control in SRE” may be the chapter that aged best. The lesson is not that reliability needs more governance. It is that teams make better reliability decisions when they can see the cost, availability, and dependency context around their own services.

Every extra nine has a price. Past a point, the next nine is paid for by customers who may never notice the last one.

Read it if you set reliability budgets. Skip it if you want a runbook. This is a policy book, not a pager book. Bulkheads before the water rises

Should this replace the favorite?👍 Promote it👎 Not ready

The Ratio is a weekly newsletter by Florian Hoeppner.
Early edition, you are one of the first.

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 👎 get replaced. Topics that get 👍 stay. No personal data is stored — just click and vote counts per section. Your feedback shapes the next edition.

Don't miss what's next. Subscribe to Reliability Economics:
← Newer THE RATIO #3: Who's watching the agents watching your systems? Older → THE RATIO #1: When the AI agent drops your prod DB
Powered by Buttondown, the easiest way to start and grow your newsletter.