**AI SRE’s Are Failing. Here’s What Actually Works.**
At a recent event in Bengaluru, I gathered a room full of seasoned SREs, platform engineers, and engineering leaders. Instead of marketing promises, they shared real stories—agent failures, stale runbooks, and systems that only worked because a few individuals held critical tribal knowledge in their heads.
The consensus? **AI’s biggest obstacle in operations isn’t model capability—it’s context.**
—
### SRE’s 4-Body Problem
Operations decisions require reasoning across four interconnected domains:
1. **Code:** What was deployed, when, and how it differs from yesterday.
2. **Infrastructure State:** What Terraform declares versus what’s actually running—cloud accounts, networks, Kubernetes, IAM, and queues.
3. **Runtime Signals:** Live metrics, logs, traces, errors, SLOs, and customer-impacting alerts.
4. **Operational Knowledge:** Post-mortems, architectural records, runbooks, and the Slack threads explaining why things are the way they are.
Each domain has tools—but no system reliably connects them. Decisions happen at the intersection, and historically, that gap was held together by “people putty”: a few overloaded experts whose institutional memory could disappear overnight.
—
### What the Discussions Confirmed
Across panels and hallway talks, a pattern emerged:
– **Agents struggle with context fragmentation.** When reasoning depends on scattered data, failures are plausible, not just incorrect.
– **Trust depends on data quality.** Autonomous root cause analysis is only as good as the graph it references.
– **Stale runbooks are dangerous.** Outdated procedures invite confident mistakes.
– **Most teams are between copilots and autopilot.** Few are fully autonomous, and few understand what it takes to climb the next rung.
Crucially, faster incident response isn’t about better dashboards—it’s cross-body correlation. Agents should aim to prepare SLO-anchored hypotheses before a human even joins the war room.
—
### The Substrate Comes Before the Agents
The room agreed: **the bottleneck isn’t model quality.**
You can’t automate reliably when agents must reconcile four disconnected systems. Fragmented context leads to hallucination—often confident and convincing, but wrong.
The foundational work is building an **active, versioned knowledge graph** that unites all four bodies and their edges:
– A commit triggers a service change.
– Terraform provisions infrastructure.
– Kubernetes rolls out the deployment.
– OpenTelemetry traces reveal latency, while Prometheus tracks SLO burn.
– The graph links these events to past incidents and proven remediations.
Only then can agents reason across code, infrastructure, runtime, and memory—and write results back for future learning.
—
### Trust Is a Decision Trace
Autonomy without auditability is unreliable. For every agent action, you need a durable record of:
– Inputs (graph snapshot version)
– Policies active at the time
– Model version
– Hypotheses considered and rejected
– Action taken and outcome
This isn’t compliance—it’s what CISOs, risk officers, and regulators will demand the first time an agent acts in production.
—
### The Path Forward
1. **Treat operations as data.** Integrate code, infrastructure state, signals, and knowledge into a single queryable graph.
2. **Embed agents in the path to production—not after it.** They must reason continuously and reduce incidents upfront, not just document them later.
Will we see fully autonomous operations in 2026? No—but the direction is clear. Scale comes from better context and trustworthy agents, not more humans.
Start with the graph. Start with the four bodies. The agents will follow—and finally, they’ll work.
—
**Original article:**
[What a room full of senior SREs confirmed about the trust gap, and where the actual work begins](https://thenewstack.io/what-a-room-full-of-senior-sres-confirmed-about-the-trust-gap-and-where-the-actual-work-begins/)
