Reliability Economics

Archives
Log in
Subscribe
July 28, 2026

AI agents can ship. Can you afford the cleanup?

The Ratio #11

Issue 11July 28, 2026

The Ratio

A weekly newsletter on reliability economics

 

The Number

2 of 101

Only 2 of 101 organizations in the benchmark spend more than half their reliability budget on prevention, the other 99 spend the majority reacting to failures after they start.

The benchmark asked organizations how much of their reliability spend goes toward prevention. 23 spend under 10%. 26 spend 10–24%. 50 spend 25–49%. Just 2 spend 50% or more.

That's 99 of 101 organizations spending most of their reliability budget reacting to failures, not stopping them.

Firefighting

Every SRE framework says shift left, invest in prevention. Almost nobody does. 2 out of 101 is not a rounding error. It's a structural reality.

This is the fire department problem in building codes. Everyone agrees prevention is cheaper than response. But fire departments get funded after the fire, not before. Prevention budgets compete with visible, urgent work, and they lose. Every sprint, every quarter.

Any team that crosses the 50% prevention mark has essentially no competition at that level of practice. Not because others can't get there. Because nobody is trying.

99 of 101 organizations spend more on cleaning up failures than preventing them.
Did this land?More like thisSwitch it up
 

This Week in Reliability

AI Agents Need Harness

From Google SRE to startup launches, the industry is converging on a single problem: AI agents can code and deploy, but without runtime harness—guardrails, observation, rollback—they shift cost from development into operational chaos. The reliability economics question is no longer whether to adopt agents, but how much you'll spend cleaning up after unsupervised ones.

 

Deep Reads

AI Engineering: Reliable Operations

sre.google  ·  Primary evidence

Google SRE published a new resource on reliable operations for AI engineering workloads, detailing practices for managing AI model deployment, inference reliability, and operational patterns specific to AI systems at scale.

Prevention

When Google SRE writes a named practices document, it's because the pattern has matured past experimentation. This isn't about training infrastructure—it's about productionizing agent workflows with the same rigor Google applies to Search. The implication: AI operations are now a distinct reliability domain with predictable failure modes and measurable prevention ROI.

Google SRE formalizes AI ops as a distinct reliability domain.

The AI Jenkins Moment

Not A Factory Anymore  ·  Historical precedent

The article argues we're at an inflection point for AI tooling similar to when Jenkins emerged—lots of custom scripts, no standard deployment pattern, teams reinventing CI/CD independently. AI agents today lack shared runtime infrastructure, observability hooks, or rollback conventions.

Neutral

This is the reliability economics framing the industry needs: every team building custom agent harness is duplicating prevention spend. The Jenkins parallel matters because it predicts consolidation—someone will ship the standard runtime, and laggards will carry double the operational cost. The question isn't if you'll adopt harness patterns, but whether you'll pay to invent them or buy them off the shelf.

Custom agent harness duplicates prevention spend across every team.

Dynatrace's new agents can reveal the single hardest part of AI operations

The New Stack  ·  Vendor response

Dynatrace announced autonomous SRE agents that integrate with its observability platform to identify and remediate issues in AI operations workflows, focusing on the complexity of distributed inference and model serving.

Firefighting

Vendor response to the harness gap—if you can't see what agents are doing in production, you can't budget for cleanup.

How AI impacts site reliability engineering

InfoWorld  ·  Adjacent signal

Article covers how AI tools (copilots, automation) are reducing toil and accelerating incident resolution for SRE teams by correlating signals across distributed systems and automating repetitive operational tasks.

Firefighting

AI for SREs vs. SRE for AI—this is the reactive half of the story, speeding up war rooms instead of preventing fires.

Why Upfront Alignment Determines Whether Complex Programs Succeed

Not A Factory Anymore  ·  Adjacent pattern

Article argues that complex engineering programs fail when teams defer hard decisions, unresolved dependencies, and untested assumptions to later phases, resulting in perpetual '80% complete' status and compounding rework cost.

Prevention

Agent deployments without upfront SLO/rollback/observability alignment will hit this wall—you can't retrofit harness into a live agent swarm.

Pilot Protocol launches to power the agent economy

The New Stack  ·  Vendor response

Pilot Protocol is a new open framework designed to standardize how AI agents interact with infrastructure, coordinate across services, and integrate with existing DevOps toolchains, positioning itself as the 'HTTP for agents.'

Prevention

This is the Jenkins contender—if Pilot Protocol gains traction, it becomes the standard harness layer and everyone else pays integration tax.

Cloudflare open-sources a debugger for privacy protocols used by Apple and Microsoft, with AI agents in mind

Cloudflare  ·  Adjacent infrastructure gap

Cloudflare released pvcli, an open-source debugger for privacy-preserving protocols (Oblivious HTTP, Private Access Tokens) that Apple and Microsoft use, explicitly designed to help developers troubleshoot agent-to-service interactions that go through privacy layers.

Neutral

Agent harness isn't just observability—it's debuggability through privacy boundaries, which nobody budgeted for in 2025.

Did these land?More like thisSwitch it up
 

The Crowd Favorite

  1. Thunderstruck — AC/DC — Synchronized reconnection after a cache flush re-saturates the recovering layer. Jitter isn't optional, it's blast radius control.
  2. Let It Be - Remastered 2009 — The Beatles — A change freeze extends deploy batch size. The first release after the freeze carries compounded risk, not reduced risk.
  3. Don't You (Forget About Me) — Simple Minds — A runbook nobody has read in 90 days has drifted. Tabletop exercises find the gaps before production does.
  4. Welcome To The Jungle — Guns N' Roses — Every undocumented upstream dependency is invisible blast radius your SLO doesn't model.
  5. Sweet Home Alabama — Lynyrd Skynyrd — A failover region that has never carried production traffic is an untested hypothesis, not a DR plan.

Prevention

Five failure modes hiding in plain sight
The playlist — keep or kill?Keep the crownDethrone it
 

The Challenger — Career Spotlight

A fintech SRE team, 8 engineers in a roughly 600-person org, opened 2023 at a 0.6 prevention-to-firefighting ratio. Eighteen months later: 2.1.

The lever wasn't headcount. It was schedule. On-call rotation now owns the SLO dashboard review in weekly planning. Prevention work gets a standing slot instead of losing to features every sprint. That alone changed backlog priorities.

Toil tracking surfaced three recurring alert categories eating 40% of response hours. They automated two. Moved the ratio more than any hiring round had.

Prevention

Schedule changed the ratio. Headcount didn't.
Should this replace the favorite?Promote itNot ready

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.

Don't miss what's next. Subscribe to Reliability Economics:
← Newer Your dashboards weren't built for agents Older → AI SRE agents ship, but are they fighting or preventing fires?
Powered by Buttondown, the easiest way to start and grow your newsletter.