Production AI · Research Series · Volume I of III
The next decade of AI will not be won by the smartest agents. It will be won by the most reliable ones.
Enterprises don't deploy prompts. They deploy systems that run for hours, act on their own, and touch money, patients, and supply chains. Agentic Reliability Engineering names the discipline those systems need — and gives it a framework, a maturity model, and seven metrics you can baseline this quarter.
44 pages · Free · By Yaron Schneider, co-creator of Dapr and KEDA
The problem
Your agents are capable. That was never the hard part.
The demo performs. Then it meets production: a model that times out, a tool that returns malformed data, an approval that takes four hours, two other agents to coordinate with. The agent loses its place. Work is silently dropped, double-executed, or completed in a way no one can reconstruct or verify.
The demo was a question of intelligence. Production is a question of reliability — and that's structural. Agentic systems are long-running, non-deterministic, and dependent on components that fail independently. For systems like that, failure is not an edge case. It is a normal operating condition.
Site Reliability Engineering doesn't cover it. SRE assumed short request lifetimes, stateless handlers, deterministic retries, and observability built around requests rather than reasoning — every one of which agents violate. And the gap widens as models improve, because ambition scales faster than the reliability substrate underneath it.
“The industry is optimizing the intelligence of agents while underestimating the reliability of the systems that contain them. That asymmetry is the defining enterprise risk of the agentic era.”
Author
Yaron Schneider, co-founder and CTO of Diagrid, co-creator of Dapr and KEDA — two CNCF projects built around the problem this report describes: making long-running, stateful workloads survive failure. Agentic systems are the next workload with that shape.
Get Volume I
44 pages, free. We'll email the PDF and send Volumes II and III when they publish.
What's inside
A discipline, not a checklist of tips.
The ARE framework — five pillars
Durability, governability, observability, verifiability, autonomous operations. Five questions to answer “yes” to before trusting an agent with real authority. Cumulative, not a menu.
A maturity model — eight levels
Prompt-based (L0) to autonomous enterprise (L7). Most enterprises run at Level 2 while funding Level 5–6 ambitions. That distance is your unmanaged risk.
Seven reliability metrics
Completion rate, recovery success, state loss, verification coverage, human escalation, agent MTTR, policy violation — each with a formula and a target range. Instrumentable today.
Eight failure modes traditional reliability misses
Runaway loops, lost state, model and tool failures, connector outages, approval bottlenecks, coordination failures, long-horizon drift. None exotic; all routine.
Case studies and a reference architecture
Banking, healthcare, and supply chain scenarios, plus an eight-layer architecture from business goals to infrastructure.
Three tools to use immediately
A pre-deployment checklist, a twelve-question executive readiness assessment, and a per-agent scorecard that gates on your weakest pillar rather than your average.
Key findings
Six conclusions the report defends with evidence.
Agent failure is a normal operating condition, not an exception.
The cloud-era reliability stack does not cover agentic execution.
Reliability, not model quality, is now the binding constraint on value.
Durable execution, governance, observability, and verification are becoming the non-negotiable floor.
Most enterprises are two to five maturity levels below their ambition.
ARE will become a named discipline — with roles, standards, and audits.