The model is not the limiting factor.
When an AI SRE gives a confident but wrong answer, the reflex is to improve the prompt, add another model, or connect another tool. That treats the agent as the failure.
But an agent can only reason over what the system preserved. If the HTTP response body is absent, the database result was reduced to duration, or the queue handoff lost its parent request, there is no chain of reasoning that recreates the truth.
Observability tools were not built for agents.
They were built to make a manageable human interface. That meant sampling high-volume traces, trimming payloads, obscuring sensitive fields broadly, aggregating metrics, and dividing the production story into separate products.
Those tradeoffs can work for dashboards. They are disastrous for an agent expected to explain exactly what happened to one user across one request.
A 200 response without its body looks successful even when it returned an empty result that caused the failure.
The trace, replay, database span, deployment, and agent run each describe a fragment without proving they belong to the same incident.
More tools create more calls, more schemas, and more plausible dead ends. Access is not the same as useful evidence.
RAG and connectors do not repair the picture.
RAG retrieves what already exists. A connector gives an agent faster access to another system. Neither can recover what was never captured, restore a sampled trace, or prove that five records describe the same production event.
Connecting five incomplete tools gives an agent five incomplete versions of the truth.
What production agents actually need.
Capture at the source
Preserve requests and responses, queries and results, sessions, queue payloads, model calls, tool calls, logs, metrics, and traces.
Connect the sequence
Carry one identity from the user through the browser, services, databases, workers, models, and infrastructure.
Redact precisely
Remove PII, PHI, credentials, and service-specific secrets without destroying the evidence required to explain the event.
Expose one record
Let Claude Code, Cursor, Codex, and other agents query the complete production trail directly.
The AI SRE is downstream.
The industry is racing to build agents that detect, investigate, and fix production failures. Those agents will improve. But their ceiling is set before the first prompt is sent: by what the software captured and whether the pieces still form one picture.
Fix the evidence first. Then let the agents reason.