| Issue 20 | September 29, 2026 |
The RatioA weekly newsletter on reliability economics | | |
|
The Number
59 vs 28
More organizations have SLOs that don't influence any decisions (59) than organizations with no SLOs at all (28).
59 organizations have SLOs that don't influence decisions. 28 have no SLOs at all. Only 16 have SLOs that actually drive action.
Neutral
Building SLOs you ignore costs you twice. Once to create them, once when your teams figure out nobody cares. It's like installing smoke detectors, removing the batteries, and leaving them on the ceiling for the fire inspector. The fixture is there. The protection is not. Of the 103 organizations in the benchmark with enough data to classify, 57% built the thing and then walked away from it.
| More organizations built SLOs they ignore than never built them at all. |
|
|
This Week in Reliability
SLOs for Agent Behavior
As AI agents move from demos to production operations, teams are applying SRE primitives—SLOs, error budgets, policy-driven guardrails—to agent reliability itself. The shift treats hallucinations and agent failures as measurable reliability problems, not R&D curiosities.
|
Grafana Labs Blog · Primary evidence Grafana Labs applied SLO methodology to their AI agents, measuring hallucinations and failures against error budgets. They found that treating agent behavior as a reliability problem—not just a quality problem—enabled the same reasoning frameworks teams use for services. Neutral This is the moment SRE thinking escapes infrastructure and colonizes agent development. If hallucinations burn error budget, suddenly product velocity and agent accuracy become the same conversation—and someone has to staff the oncall rotation for the agent itself. | Hallucinations as error budget: agent reliability becomes staffable. |
|
CNCF Blog · Infrastructure precedent CNCF argues that coding agents became useful when they gained capable tools, shared repositories, subagents, and skill capture—not when they got better models. The post advocates for a cloud-native harness to manage agent lifecycles and capabilities. Prevention The harness is the new container runtime: isolation, resource limits, and observability for non-deterministic workloads. This frames agent operations as an infrastructure problem, which means it inherits the entire SRE playbook—and its cost structure. | Agent harness: the runtime for non-deterministic workloads. |
|
Anyshift.io · Vendor effectiveness claim MotherDuck's Hilary Brennan-Marquez reports an agent-assisted investigation completed in 30 seconds that she estimated would take a human 20 minutes to two hours. Firefighting This is the MTTR compression case every vendor wants to tell—but notice the 20-minute lower bound, which means the agent didn't eliminate context-gathering, just parallelized it. | 30-second investigation: parallelized retrieval, not eliminated context. |
|
Anyshift.io · Adjacent: agent grounding infrastructure Two customer investigations show how a context graph was necessary for agents to retrieve the right data by tracing changes and shared dependencies, beyond what Model Context Protocol alone provided. Prevention MCPs give agents access; context graphs give them judgment about what to fetch—the difference between grep and root cause. | Context graphs: the judgment layer MCP doesn't provide. |
|
Gremlin · Adjacent: agent limitation analysis Gremlin catalogs the observability gaps AI SREs face—missing dependencies, untracked config drift, silent failures in third-party services—and argues these gaps limit agent effectiveness more than model capability. Neutral The agent is only as good as the data you didn't instrument, which means the agent tax is paid in observability coverage you should have bought anyway. | Agent effectiveness ceiling: your observability coverage floor. |
|
Densify (KubeX.ai) · Adjacent: policy precedent Article surveys platforms offering autoscaling and policy-driven resource management, emphasizing that policies define boundaries autoscaling operates within—who can use how much, which workloads can change, and what safeguards must be respected. Prevention Policy-driven resource management is the pattern agents will need: bounded autonomy with audit trails, which is exactly what SLOs and error budgets provide for reliability. | Autoscaling policies: the template for agent guardrails. |
|
ilert · Adjacent: regulatory forcing function The Digital Omnibus moved EU AI Act high-risk compliance to December 2027. Article 73 requires serious incident reporting but doesn't mandate a 72-hour window. Ilert explains what still applies in 2026 and how to build reporting now. Neutral If agents are high-risk systems under the Act, then incident reporting becomes a compliance artifact—which means your agent's error budget is now an audit document. | AI Act compliance: error budgets as audit artifacts. |
|
|
The Crowd Favorite
- Wonderwall - Remastered — Oasis — The dependency you never mapped is the one that owns your next SEV-1. No owner assigned.
- 99 Problems — JAY-Z — A 99% false-positive rate in your alert queue guarantees the real P0 arrives unnoticed.
- Enter Sandman (Remastered) — Metallica — Silence from a service is not health. Deadlocks and OOM kills need their own alerts, not just HTTP checks.
- Killing In The Name — Rage Against The Machine — Undocumented manual overrides survive every runbook revision. They detonate on the next automated rollout.
- Waterloo — ABBA — Every cascade has one point where load shed would have contained it. The post-mortem exists to find that point before the replay.
Prevention
| Map the failure before it pages you |
| The playlist — keep or kill? | |
|
|
The Challenger — Tool of the Week
incident.io centralizes incident declaration, role assignment, and retrospective generation inside Slack. This is the reaction side of the U-Curve. It doesn't reduce how often things break. It cuts the coordination tax during and after each fire.
Honest take: under 20 engineers, a shared channel and a Google Doc cover 80% of this for free. Worth paying for when war-room coordination itself becomes the bottleneck, which is the failure mode this tool was actually built to solve. Not the fire itself.
Firefighting
| Faster cleanup, not fewer fires |
| Should this replace the favorite? | |
|
|
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.
|
|
|