Agents join the on-call rotation, with humans still holding the call

An agent that starts the investigation

Dynatrace announced on 11 November that its SRE Agent is generally available. The company says the agent triages incidents based on severity, impacted users, and blast radius, and analyses incidents in minutes rather than hours. The company says the agent is embedded as an action within a workflow, follows Smartscape dependencies upstream and downstream during investigations, and integrates with Slack, Microsoft Teams, email, JIRA, and ServiceNow. Dynatrace Intelligence, the company says, groups events with the same cause to prevent alert floods. m:…9129

Dynatrace also says the agent is built to keep the decision with the person. The company says the SRE Agent operates on a model where the user retains decision-making authority, and that users can build custom agents using the Prompt agentic AI action. Beyond incident management, the company says the agent can generate right-sizing analysis for cloud resources, forecast capacity limits for clusters or disks, and identify production systems exposed to new security threats. Those capabilities are the company's claims about its own product, and they are reported as such. m:…9129

Resilience is being rebuilt around the agent

In a second post on 11 November, Dynatrace describes how it is changing its own development process for an age of AI-operated systems. The company says it is integrating resilience patterns such as backoff, jitter, and circuit breakers into its AI software development lifecycle, and strengthening its QA process by incorporating chaos engineering and fault-injection exercises. The company says it is upgrading circuit breakers and request-volume controls to strengthen fault isolation across shared services, and has implemented safepoints in the telemetry ingest pipeline to protect processed data from dependent service failures. m:…9130

Dynatrace gives dates for the work. The company says it targets core changes to fault isolation for December 2026, plans to implement regional isolation by the first half of 2027, and targets further telemetry ingest decoupling for the first half of 2027. The company also says it is implementing an enhanced SRE model starting December 2026 to improve incident response and recovery, and is advancing cross-region return-to-operations capabilities using an active-active deployment model. Those dates are the company's stated plans, not completed work. m:…9130

Human in the loop is not a safety plan

Rootly, also on 11 November, pushes back on the idea that a human watching the agent is enough. The authors argue that human in the loop is a repackaging of the outdated pay attention directive, and that it is often invoked to pass the buck for poor system design onto the user. They argue that human attention is not a quantifiable resource that can be extended, and that automation bias makes humans particularly ineffective at human-in-the-loop oversight when AI is involved. Their point is that attention cannot be treated as a control that scales with the number of agents. m:…4597

Rootly's alternative, as described in the post, is to change the system rather than the watcher. The authors say redesigning systems is more resource-intensive than relying on human operators, but argue for it anyway, and say continuous verification frameworks can prove system output meets desired parameters. The post also says responsibility for AI agents is currently an intractable problem. That is an argument from Rootly's authors, not a measured result, and it sits in tension with Dynatrace's human-approved model: one side keeps a person in the decision, the other says the person cannot reliably catch what the agent gets wrong. m:…4597

Credential attacks start with a valid login

Databricks, writing on 11 November, describes a different on-call job: spotting attacks that begin inside the perimeter. The company says credential-driven attacks often begin with a valid login rather than a perimeter breach, and that adversaries use non-linear attack sequencing to map environments and retrieve new credentials. Databricks says five or more distinct discovery API calls within one minute is a potential indicator of non-human reconnaissance, and that Genie includes over 30 built-in detections for audit logging. The company says its audit logs record events in a service_name.action_name format, and that verbose audit logging is required to capture command text in commandSubmit events. m:…0729

The response Databricks describes is thorough and unforgiving. The company says deactivating a user at the account level blocks access across all workspaces, but that OAuth tokens minted by a service principal remain valid for up to one hour after deactivation. Databricks recommends rotating every credential an identity could have reached during a suspected compromise, and says adversaries may use serverless compute as a fallback execution path after being blocked. Oracle, also on 11 November, describes a parallel effort: the company says it is developing AI-assisted application security that connects infrastructure and application signals with business context, through three layers of signals and context, reasoning and validation, and engineering action, with specialised agents investigating vulnerabilities and gathering corroborating evidence. Oracle says investigation results are retained in a persistent knowledge store for recursive learning, and that remediation is verified to ensure prohibited behaviour is restricted while business workflows function. m:…0729 m:…2583