Observe

AI SRE reaches outside systems for context rather than to act in them. Each user connects Notion and Atlassian through those services' MCP servers so the agent can pull runbooks, pages and tickets into an investigation, and inside Observe it searches datasets, runs OPAL queries and inspects APM service maps under the user's own permissions. Findings come back as recommendations in the chat; the agent does not file tickets, page anyone or change infrastructure, and alert notifications to Slack, PagerDuty, Teams or email are monitor actions rather than agent actions.

Sourcedocs.observeinc.com/docs/configure-ai-sre-connectors.mdread 2026-11-09

A single agent plans each investigation as a chain of tool calls: resolve correlation tags from the question, find the datasets that carry them, generate an OPAL query and run it, then inspect service maps or traces and summarize. Skills turn a repeatable playbook, such as debugging frequent pod crashes, into instructions the agent follows step by step when someone types its slash command. The trace and monitor assistants are the same agent with page context, and nothing coordinates several agents or runs a branching workflow around them.

Sourcedocs.observeinc.com/docs/model-context-protocol.mdread 2026-11-09

Investigations draw on the customer's telemetry at run time, guided by a context graph in which Cortex Search covers correlation tags and Cortex embeddings cover dataset metadata, so the agent can map a phrase like a container name to the right datasets. Notion and Atlassian supply runbooks and tickets through connectors when a user has linked them, read per investigation.

Admins can also upload runbook files on an Uploaded documents page, but those ground o11y AI Help, the separate how to assistant for questions about Observe, rather than AI SRE, and the indexes AI SRE holds describe where data lives rather than the team's operational knowledge.

Sourcedocs.observeinc.com/docs/where-does-your-data-go.mdread 2026-11-09

Guardrails here limit what the agent can see and how often it runs rather than gating actions, because it recommends instead of acting. Queries run with the user's own dataset permissions, RBAC decides who may use AI SRE or create shared skills, and an OAuth MCP session can be narrowed to read only. Background investigations use read only tools, run as the user who last saved the monitor, check that identity before each run, and stop at ten per monitor and 1,000 per account a day, with muted alerts skipped. No approval step exists for changes because none are made.

Sourcedocs.observeinc.com/docs/auto-investigate-alerts-with-ai-sre.mdread 2026-11-09

Access runs through groups and members with RBAC privileges, a permission manager and dataset query filters, single sign on with Microsoft Entra ID, ADFS, Google Workspace, Okta, OneLogin or PingOne, local authentication policies, service accounts managed in the UI or by API, and an audit trail of user activity.

AI SRE and the MCP server inherit the caller's permissions with no escalation, and an OAuth MCP session can carry narrower rights than the user. The compliance reports behind the product are Snowflake's and cover Snowflake's own platform, while Observe's own trust center is closed to visitors, so no current certificate names Observe itself.

Sourcedocs.observeinc.com/docs/ai-sre-permissions-and-access.mdread 2026-11-09

Every investigation is saved as a chat holding the agent's messages and query cards, which readers rerun under their own permissions, and alert triggered runs are linked from the alert with a status and a recorded reason whenever one was skipped. Chats can be shared or copied to continue under another identity, with access copied from the monitor. An audit trail tracks user activity, though actions taken by a background investigation appear under the user who last saved the monitor. LLM Explorer traces the customer's own LLM workflows step by step.

Sourcedocs.observeinc.com/docs/auto-investigate-alerts-with-ai-sre.mdread 2026-11-09

AI SRE keeps persistent memories per user across chat sessions: short facts it writes during or after a conversation, such as last week's root cause on a server, a path already ruled out, or that the gateway means the payment gateway for this person, and retrieves by semantic search in later sessions. Users can ask it to remember something and can switch memories off in their settings. Memories are on by default, and no retention period or way to review and delete individual memories is given.

Sourcedocs.observeinc.com/docs/ai-sre-persistent-memories.mdread 2026-11-09

Each tenant lives in one of eight independent regions with no data shared between them: AWS US West in Oregon, US East in Virginia, Canada Central, Frankfurt, the United Kingdom, Tokyo and Sydney, or Google Cloud US Central in Iowa, where telemetry passes through AWS Oregon before warehousing. A customer can hold tenants in more than one region. AI SRE calls Snowflake Cortex models, and some requests route across regions to reach an available model.

Sourcedocs.observeinc.com/docs/regions.mdread 2026-11-09

Teams mostly build their own playbooks as skills: a name, a description and Markdown instructions, private or shared with the workspace, created from the settings page, by asking the agent to save a conversation as a skill, or through the observe_skill Terraform resource. Observe also supplies some skills of its own, and the home page offers preloaded questions. Data onboarding skills for coding agents install the Observe Agent and OpenTelemetry instrumentation in a public preview, and the trace and monitor assistants are one agent working with page context.

Sourcedocs.observeinc.com/docs/create-skills-for-ai-sre.mdread 2026-11-09

Generally available entry points are people asking: the AI SRE chat on the home page, the assistants inside Trace Explorer and monitor creation, and outside agents calling the MCP server.

Alert triggered investigation exists as an invitation only private preview, enabled per tenant and limited to internal evaluation; once switched on for a monitor it investigates new alerts of chosen severities in the background, ten per monitor per day by default, and links the findings from the alert and its notification. Opened to every customer for production use, that would put work on the agent from events.

Sourcedocs.observeinc.com/docs/auto-investigate-alerts-with-ai-sre.mdread 2026-11-09

AI SRE runs on models hosted in Snowflake Cortex, which serves as its model gateway, with Cortex Embed for memories and Cortex Search for documentation. Observe picks the models, names no model or maker behind the completions, and offers customers no setting to change them. Coding agents that connect through the MCP server use whatever model their own tool runs, which leaves the choice with that tool rather than with Observe.

Sourcedocs.observeinc.com/docs/ai-sreread 2026-11-09

Outside callers drive Observe through its developer API, including service accounts and reference tables managed over HTTP, and a Terraform provider manages monitors, RBAC settings, credit controls and AI SRE skills as code. An MCP server at each tenant takes an OAuth login or an API token and exposes investigation tools to Cursor, Claude Code, Claude Desktop and Cortex Code, with a plugin in the Cursor marketplace and a CLI carrying many of the same tools.

Sourcedocs.observeinc.com/docs/model-context-protocol.mdread 2026-11-09

LLM Explorer gives teams step by step traces of their own LLM applications through OpenLLMetry auto instrumentation, covering OpenAI, Anthropic, Bedrock, Gemini, Vertex, Mistral, Cohere, Groq, Ollama and others, which helps them troubleshoot after the fact. It records rather than scores, with no test sets, judges or quality metrics. AI SRE's own answers are checked by reading the query cards behind them, and nothing measures its findings against a known result.

Sourcedocs.observeinc.com/docs/llm-observability.mdread 2026-11-09

The agent never drives a browser or desktop. It works through dataset searches, OPAL queries, APM tools and MCP connectors to Notion and Atlassian, and its assistants live inside Observe's own Trace Explorer and monitor pages, where the product itself runs in the browser. Data onboarding skills run inside the customer's coding agent rather than as an Observe agent operating an interface.

Sourcedocs.observeinc.com/docs/ai-sreread 2026-11-09