Dynatrace

Workflow steps act in outside systems through connectors installed from Dynatrace Hub.

Jira issues are created, transitioned, commented and assigned, ServiceNow receives incidents and ITOM events, PagerDuty pages on call staff, Red Hat Ansible runs jobs and Jenkins starts builds, while the Kubernetes, AWS and Azure connectors change pods, deployments, EC2 instances, autoscaling groups and virtual machines.

Slack, Microsoft Teams and Microsoft 365 carry messages, and GitHub, GitLab, Snowflake and Microsoft Entra ID round out the catalog. The structured output of an agent step decides which of these actions runs next in the same workflow.

Sourcedocs.dynatrace.com/docs/analyze-explore-automate/workflows/default-workflow-actions/actionsread 2026-11-09

Dynatrace Workflows is the runtime for the agents. A workflow chains a trigger, DQL queries, code actions, connector actions and one or more agent steps, and later steps read the output of a Prompt agentic AI step field by field to decide what happens next.

Run agent drops a ready made agent into a step with no prompt to write, and the platform coordinates several Dynatrace agents on one multistep task, so an investigation can feed a ticket and a remediation job in the same run.

The agentic actions still carry a preview label and run in any environment on a platform subscription with AI units, up to 1,000 runs an hour per workflow.

Sourcedocs.dynatrace.com/docs/dynatrace-intelligence/dynatrace-intelligence-integrations/agentic-workflowsread 2026-11-09

Each run draws on the customer's own environment in Grail: logs, metrics, traces, events, deployment events, security findings and the Smartscape map of how hosts, services and processes depend on each other, reached through DQL and the tools picked for the step.

A document search tool finds notebooks, dashboards and troubleshooting guides kept in the environment. Agentic Assist adds a curated set of skills and reference files that Dynatrace maintains on observability practice, and generative answers lean on Dynatrace's own documentation, community and hub, so context is gathered fresh for each investigation.

Sourcedocs.dynatrace.com/docs/dynatrace-intelligence/agentic-and-generative-ai/chat-with-dynatrace-assistread 2026-11-09

A Request Approval step pauses any workflow until named users or groups approve or decline by email, and the Slack connector can request the same approval in a channel; a decline or a timeout fails the run, so a remediation waits for a person before it lands.

Each agent step can call only the tools picked for it, up to 20, under the permissions of the user or service account that runs it. PII blocking stops agentic prompts carrying card numbers, social security numbers, IBANs or email addresses, with patterns chosen per environment, and content filters screen prompt injection and off topic requests.

Sourcedocs.dynatrace.com/docs/analyze-explore-automate/workflows/default-workflow-actions/approval-request-workflow-actionread 2026-11-09

Users and groups get granular permissions, down to who may run agentic AI, invoke the MCP server, run or edit workflows and read each storage table, and they sign in through SAML single sign on with IP allow lists available. Platform tokens and OAuth clients carry only the scopes granted to them, and an OAuth client's token never exceeds its creator's own rights. Dynatrace holds ISO 27001, SOC 2 Type II, FedRAMP Moderate and HIPAA, backed by annual third party audits, penetration tests and red team assessments.

Sourcedocs.dynatrace.com/docs/manage/data-privacy-and-security/data-security/data-security-controlsread 2026-11-09

Every workflow run lands in an execution monitor that records each step of an agent's run along with the decisions on any approval steps, and the SRE Agent labels its own annotations on the problem so readers can tell agent findings from people's notes. Security relevant events, such as configuration changes and access to the environment, go to audit logs the customer queries in Grail with DQL. Support staff access to an environment is role based, needs multifactor sign in from outside the corporate network and is itself logged.

Sourcedocs.dynatrace.com/docs/manage/data-privacy-and-security/data-security/data-security-controlsread 2026-11-09

Dynatrace Assist saves every conversation automatically, and a user can reopen, rename or delete it from the conversation list. In agentic mode each person can also record a job title, experience level and preferred format, which shape the style and depth of later answers. Agents in workflows leave their findings on the problem as labeled annotations, which belong to the incident record rather than to a store the agent reads back, and each investigation starts from the telemetry rather than from facts kept by earlier runs.

Sourcedocs.dynatrace.com/docs/dynatrace-intelligence/agentic-and-generative-ai/chat-with-dynatrace-assistread 2026-11-09

Each Dynatrace environment lives in a region the customer picks across three clouds. AWS offers US East, US West, Ireland and Sydney, with London, Frankfurt, Canada, São Paulo, Singapore, Mumbai, Tokyo and Tel Aviv on request; Azure offers Virginia, Arizona, the Netherlands, Toronto, Dubai, Zurich and Sydney on request; Google Cloud offers Virginia and Frankfurt on request. Platform data for each environment sits in its own storage with its own encryption key, and prompts sent to the agents' language models are processed in an EU region for EMEA environments and a US region for all others.

Sourcedocs.dynatrace.com/docs/manage/data-privacy-and-security/data-security/data-security-controlsread 2026-11-09

Ready made agents ship as workflow templates a team creates in one step and then adapts: the SRE Agent, Alert Reduction, Database Operations, Infrastructure Optimization, Kubernetes Troubleshooting, Kubernetes Operations and Mobile Crash Remediation for operations, and Threat Triage, Vulnerability Verification, Security Insights Report and Security Association for security teams. Each does its own job; the SRE Agent, for one, turns a new problem into a root cause, findings and a recommended fix written onto the problem, and the Run agent action drops a prebuilt agent into any custom workflow.

Sourcedocs.dynatrace.com/docs/dynatrace-intelligence/dynatrace-intelligence-integrations/sre-agentread 2026-11-09

Work reaches the agents through workflow triggers. A problem trigger fires when Dynatrace opens a problem that matches a filter on state, category, severity, minimum duration, affected entities and tags, and by default waits until the initial root cause analysis has finished. Schedules run agents on recurring work, such as a weekly review of where alert noise comes from, and event triggers react to other events in the environment. People also start work from Dynatrace Assist and from conversation starters embedded in Dynatrace apps, and outside agents call in over the remote MCP server.

Sourcedocs.dynatrace.com/docs/dynatrace-intelligence/dynatrace-intelligence-integrations/sre-agentread 2026-11-09

Dynatrace picks the models behind its agents. Prompts go to large language models hosted on Microsoft Azure AI and AWS Bedrock under enterprise terms that bar the hosts from storing the data or training on it, and a DQL generation model that Dynatrace fine tunes on its own employees' queries turns questions into queries. Customers cannot choose a model or provider for Assist, the ready made agents or the workflow actions. Outside agents that connect through the remote MCP server keep whatever model they already run.

Sourcedocs.dynatrace.com/docs/dynatrace-intelligence/agentic-and-generative-ai/agentic-and-generative-ai-data-privacyread 2026-11-09

Outside callers drive Dynatrace through its REST APIs with platform tokens or OAuth clients, governed by the same permissions as the console. A remote MCP server at each environment's own address accepts bearer tokens or an OAuth authorization code flow and offers tools to generate, explain and run DQL, investigate problems and vulnerabilities, analyze Kubernetes events, forecast time series, find documents and resolve entity names. Ready plugins connect Claude Code and Cursor, and calls made this way draw AI units like work done inside workflows.

Sourcedocs.dynatrace.com/docs/dynatrace-intelligence/dynatrace-mcpread 2026-11-09

Checking an agent change is a manual review. To validate the SRE Agent, a team replays or creates a synthetic problem that matches the trigger, reads the resulting annotation for accuracy and evidence, then widens the lookback window or trace sample in the prompt if it falls short.

Thumbs up and down ratings in Assist go to Dynatrace's product team rather than into a score the customer sees, and no scored test cases run against an agent before a change goes live. Site Reliability Guardian scores the customer's own releases against service objectives, which tests the services rather than the agents.

Sourcedocs.dynatrace.com/docs/dynatrace-intelligence/dynatrace-intelligence-integrations/sre-agentread 2026-11-09

No agent here drives a browser or a desktop. Synthetic browser monitors replay recorded clickpaths against the customer's own sites from public or private locations, and a workflow can run one on demand, but that is a scripted monitoring check rather than an agent navigating an interface. The agents reach systems through DQL, MCP tools and workflow connectors, and Assist reads what is on screen in Dynatrace's own apps only to describe or summarize it.

Sourcedynatrace.com/pricingread 2026-11-09