Autopilot release notesRSS

Bug fixes

• Faster, more resilient chat panel: The chat now opens faster, queues bursts of requests, and automatically retries calls that fail due to a brief network hiccup instead of initially surfacing an error.
• Clearer messaging when you hit a usage limit: Instead of a generic error, Autopilot now tells you specifically when your hourly request limit or daily usage limit has been reached.
• More reliable entity search: Entity lookups now correctly perform fuzzy matching and are more likely to prefer entities in the account you're asking about, so you're more likely to find the right one on the first try.
• Fixed incorrect time windows: "IST" could previously be misread as Israel Standard Time instead of India Standard Time, leading to an analysis window that was off by several hours or an unexpected validation error. IST now resolves correctly, along with several other natural-language time-parsing fixes.
• More accurate chart results: Valid chart values are no longer stripped from a visualization when it shouldn't have been.
• Coverage-gap questions now have a more complete answer: Fixed an issue where asking about monitoring coverage gaps (for example, missing alert or entity coverage) might not actually use Autopilot's dedicated coverage-gap analysis, leading to a less complete answer.
• Fewer failed requests when attached context is incomplete: Fixed an issue where a request with incomplete attached context could fail outright instead of falling back to your original question.

New features

• Deeper root cause analysis when you're paged: Autopilot can now help trace an alert back to the code change behind it from more entry points, including incident and Slack-triggered investigations, so you can pinpoint the issue without digging through the entire deployment history.
• See the full data, not just the chart: Tool results now show the underlying data alongside a visualization, so you can more easily verify what Autopilot based its answer on.
• Spot intermittent spikes, not just averages: Metric analysis now surfaces percentile detail and trend history, making a brief spike easier to catch than it would be from an average alone.
• Ask about a full month of alerts in one request: The alert lookback window for entity-scoped queries has been widened from 7 to 30 days, and an oversized request now gets automatically scaled back instead of rejected outright.
• Jump straight into your session: Responses now include direct session links for browser and mobile, so you can pick up where Autopilot left off without hunting for it.

Improvements and enhancements

GitHub MCP Integration Capability Updates (Public Preview)

Autopilot's GitHub integration now ties active alerts more directly to your deployment and release history, so you can triage without switching between tools.

• See in-flight fixes before duplicating the work: During alert triage, Autopilot now checks GitHub for open pull requests already targeting the affected service, so you can see whether a hotfix or related fix is already underway before starting your own.
• Spot version drift the moment an alert fires: Autopilot now compares the release actually deployed against the version that was running when the incident started, surfacing recent rollouts or unverified deployments that may be contributing to the issue.

Bug fixes

• Fixed queries that could fail when counting unique values: Questions that required counting something unique — like how many distinct users hit a specific error — could cause Autopilot to build an invalid NRQL query and fail to respond. Autopilot is now guided to better handle this, making these questions less likely to error out.
• Fewer rejected error-analysis responses: Fixed an issue where asking for a detailed error breakdown could get rejected for being too large to deliver.
• Fewer timeouts analyzing high-volume logs: Increased the time allotted for log analysis so accounts with large log volumes are less likely to time out.
• Cancel to stop the run: Canceling an investigation from the chat panel now stops it running on the backend.
• Faster, cleaner chat panel load: The chat panel now loads Ask AI and Autopilot first instead of waiting on your full agent list, and no longer shows either one twice in the agent dropdown.
• More consistent knowledge-base grounding: Fast, direct-answer responses and alert-triage investigations now draw more consistently on your knowledge base before answering.
• More accurate deployment status in GitHub investigations: Better reporting of service release status.
• Clearer guidance when Autopilot can't find an answer: Replaced a generic "something went wrong" message with more helpful guidance to try rephrasing your question.
• Corrected a misleading connection error: A login-related redirect was being reported as an expired connection; it's now identified correctly.

New features

• Autopilot now applies improved, tailored size limits and smarter budget tracking to every major tool, from NRQL queries to APM and alert analysis, so one oversized response or a single failing tool call is less likely to impact a whole investigation.
• Behind-the-scenes improvements to how Autopilot compresses context, routes requests, and retrieves memory and knowledge help cut processing time without sacrificing answer quality, especially on high-volume accounts. Average run time is down about 20%.
• Improvements for investigating complex environments with many entities, alerts, or logs, where the sheer volume of data could previously cause the run to fail.

Bug fixes

• Autopilot now better grounds Slack responses to the data it actually retrieved.
• Autopilot now better recovers from individual step failures and continues where it can, resulting in fewer incomplete analyses.
• Fixed an issue where a complex query could cause Autopilot to fail silently. These cases now fail fast with a clear explanation so you can make adjustments.
• Generic "something went wrong" messages have been replaced with suggestions that provide specific detail about what happened and what to try next.
• The chart for "how many active issues" now better matches the text details in the response.

What's new

Autopilot: Faster, More Reliable Investigations

• Faster investigations: Autopilot now dispatches steps in parallel, so you see quicker turnaround and fewer timeouts, especially on requests that fan out to multiple sub-investigations.
• Fast Path Routing: You'll get instant answers on simple requests. A new fast-path engine recognizes straightforward questions and skips unnecessary steps, so you see a response faster.
• Save memories in thread: Tell Autopilot in chat to save information to memory and it will automatically add that context to long term memory.
• One-Click NRQL Copy: Copy generated NRQL queries directly to your clipboard with a single click from the chat panel.
• One-click chat expansion: Instantly maximize the chat panel with a single click for a larger, uncluttered workspace.
• Query navigation shortcuts: Cycle through and re-select past queries using the Up (↑) and Down (↓) arrow keys in the prompt box.
• More reliable responses: We fixed several issues to improve investigations and give more complete answers. You'll now get a clear message if you've hit your hourly usage limit instead of a raw error, and we fixed a rare issue where an investigation could appear finished before it actually completed.

GitHub MCP Integration Capability Updates (Public Preview)

GitHub Enterprise Cloud now supported: The GitHub MCP integration now works with GitHub Enterprise Cloud in addition to public GitHub. If your organization uses a YOURSUBDOMAIN.ghe.com instance, connect it to Autopilot using the same PAT flow. The only new step is entering your GHE subdomain URL during setup. Both investigation use cases work on GHE.com: change-triggered root cause identification and manual PR/commit lookup.

Supported GitHub environments:

Setup change for GHE.com users: When connecting GitHub in Configure > Autopilot > External MCP connections, enter YOURSUBDOMAIN.ghe.com in the GitHub URL field instead of github.com. PAT permissions are unchanged: Pull requests: Read, Contents: Read, Metadata: Read. If your GHE.com organization enforces SAML SSO, authorize the PAT for SSO in GitHub before entering it in New Relic — an unauthorized token fails with 403 on first use.

Scope your PAT to repos you actively track: Autopilot reads PR descriptions and file change metadata from every repository the token authorizes.

What's new

Autopilot: More Reliable Analysis

• More accurate anomaly detection: Transaction dependency analysis now uses a more rigorous threshold. The previous setting was overly-sensitive and flagged noise as anomalies. Downstream dependency analysis also more clearly differentiates between correlation and causation, reducing speculative root-cause attribution.
• Trend comparisons grounded in real data: When Autopilot compares current error rates or counts to a prior period, those baselines now come from verified NRQL queries.
• Longer investigations complete reliably: Resolved a bug where some investigation flows did not properly perform context window compaction when they needed to occur. If compaction encounters an error, your investigation progress is preserved rather than lost. Investigations also properly stop when you navigate away, freeing browser resources.
• Correct active issue counts: Added fixes to return more accurate number of issues when asking for number of active issues in your account.
• Slack responses render cleanly: Markdown tables and structured formatting (headers, dividers, and lists) now display correctly in Slack notifications and workflow responses.
• EU entity resolution fixed: Resolved a bug that made it difficult for Autopilot running in the EU to find the Kubernetes agent.
• Your timezone and context preserved: Your timezone and the entity context you're investigating are now correctly maintained throughout the entire session. Previously, context could be lost when multiple context sources were present.
• General fixes to UI rendering when using comparison operators and special characters: NRQL queries containing comparison operators or other special characters are no longer mangled in the response visualization.
• Improved prompt-injection identification, with fewer false positives: Autopilot no longer incorrectly declines benign in-chat notes (such as providing context or clarifying details) with a misleading "prompt injection" warning. The guardrail is now properly scoped so valid requests proceed normally.
• Memory recall improvements: Fixed bugs that prevented Autopilot from finding facts from multi-memory lookups and improper flags on valid memories.
• Preferences honored correctly: Autopilot no longer incorrectly tells you it can't remember your stated preferences (like how you'd like to be addressed). Your conversational preferences are now recognized and respected.
• Documentation citations reference the right account: When Autopilot cites documentation or entity information from RAG lookups, it now references entity GUIDs belonging to your account.
• Accurate timeseries results reporting: TIMESERIES NRQL queries that return valid data are no longer incorrectly reported as having no results.
• No more repeated reasoning text: Fixed a bug where the same reasoning text could appear multiple times in Autopilot's thinking display during an investigation.
• Improved stability of the chat panel in the browser: Fixed a memory leak that could cause browser tab crashes when loading chat history on accounts with many conversations. Also fixed crashes when multiple browser tabs had the AI chat panel open simultaneously, and prevented concurrent agent streams from conflicting when starting a new conversation mid-stream.
• Better visualization summaries: Investigation responses now capture both NRQL queries and their resulting visualizations more completely, giving you richer context in summary cards.
• Incident investigation follows related issues: When investigating major incidents with related issues across different accounts, Autopilot now resolves each related issue using its own account ID rather than defaulting to the parent incident's account.
• UI polish: Feedback ratings now only persist when you explicitly submit them. Action icons are always visible (no longer hidden behind hover). Copy conversation is properly disabled until there's content to copy. Slash command connector names display correctly.

GitHub MCP Integration Capability (Public Preview)

When an alert points at a deployment-related regression, the first question is almost always "what shipped?" You end up switching between the New Relic incident, your Change Tracking event, and the GitHub PR to piece the story together. Connect Autopilot to GitHub so it can pull the Pull Request behind a Change Tracking deployment and return the "what shipped" story inline.

• Link a deployment to a PR: When Autopilot identifies a Change Tracking event with a commit SHA during an investigation, it fetches the associated Pull Request and returns the number, title, description, author, and URL as part of the investigation response. No tab switching needed.
• Summarize what changed: Autopilot reports the files touched and the added and deleted line counts for the PR that shipped the incident-adjacent deployment.
• Follow-up questions in the same thread: Once a PR is on the record, you can ask Autopilot to compare it to the previous deployment or correlate the change with the failing signal. The GitHub context stays live in that investigation.
• Works from any entry point: GitHub context flows into investigations triggered by an SRE alert destination, a Slack message, or the AI Chat panel. Autopilot brings in the PR automatically whenever a Change Tracking event with a commit SHA is present.
• Read-only access via Personal Access Token: Autopilot uses your configured GitHub PAT scoped to Pull requests: Read, Contents: Read, and Metadata: Read. It never creates PRs, posts comments, merges branches, or stores source code. This preview supports public github.com only. GitHub Enterprise Server (self-hosted) and GitHub Enterprise Cloud are not supported yet.
• Org-level setup: A New Relic organization manager connects GitHub once from Configure → Autopilot → External MCP connections. The PAT is encrypted at rest in the New Relic Secrets Service, isolated to your organization, and retrieved into request-scoped memory only for the duration of a single investigation.

Prerequisites: Active Autopilot setup, Change Tracking configured with commit SHAs flowing into New Relic, GitHub organization admin access to create a fine-grained PAT, and New Relic org manager access.

What's new

Autopilot 2.0: Remember What Matters, Investigate in Slack, and Scale to Every Account

• Memories: give Autopilot lasting context

  • Autopilot now has the option to build a persistent knowledge base from your investigations, so you do not need to re-explain the same environment quirks, workarounds, and past incidents every session.

    • Save any Autopilot response to memory with a new button in the chat panel, memories can be saved that are specific to you, your account, or your organization, so useful findings persist instead of getting lost in chat history.
    • Autopilot recalls what it learned last time: if a past investigation determined that a recurring spike is expected behavior from a scheduled job, Autopilot applies that context to future analyses without requiring you to repeat it.
    • You control what's remembered: authorized users can view, search, and delete individual memory entries from the Memory Management UI at any time.
    • Access follows your role: organization managers can manage memory at the organization, account, and user level; other roles manage their own private memory and can view what's shared at broader scopes.

  • For more information about specific functionality and controls, see Autopilot Memories.

• Get Autopilot Investigations in Slack with the Slack Integration

  • Stay in the Slack thread your team is already in: Mention the New Relic app in any channel you add Autopilot and Autopilot investigates right in the thread, so that you don’t have to bounce between Slack and the New Relic UI to triage an alert.
  • Triage alerts without retyping context: When New Relic posts an alert notification to a channel, replying with @New Relic is enough. The agent reads the notification to help identify the entity and the issue, so the first response remains focused on the target issue.
  • Keep the conversation going: Multi-turn investigations work in-thread, follow-up questions inherit the prior context, and every reply runs with the asking user's New Relic permissions.
  • Connect once, then move on: Linking your Slack identity to New Relic is a one-time step per workspace, and switching the New Relic organization tied to your Slack identity is a single click in the app's Home tab.
  • For more information about specific functionality and controls, see Autopilot Slack integration.

• Other new features

  • Turn Autopilot on or off with one click. Organization managers can now easily turn Autopilot on or off for their organization using New Relic Feature Control.
  • Connect more accounts. Now Autopilot can correlate signals across more of your accounts instead of a subset of it.
  • Attach specific knowledge documents to a prompt with a slash command. Type the slash command in the Autopilot side panel to pull a specific runbook, retro, or postmortem into the prompt. Attached documents stay in scope for follow-up prompts until you remove them or end the session.
  • Autopilot is now the default agent in the chat panel, so you land in it automatically instead of having to switch to it.
  • Autopilot is available on HIPAA-designated accounts.
  • A redesigned investigation panel is rolling out, with a new header and homepage, inline context management, a reorganized history view, a fullscreen mode, and consistent icon and author styling throughout the conversation.
  • Configuration troubleshooting. Instead of a generic error message, you may now see a specific reason and a link to fix potential configuration issues, whether that's a missing entitlement, an account that isn't configured, or a required setup step.
  • Downstream-dependency root cause analysis: investigations can now help trace an issue to the services downstream of it, for accounts with Transactions 360 enabled.

• Improvements and fixes

  • URL decoding: Fixed a class of failures where the agent couldn't determine the account ID from certain New Relic alert URL formats (including HURL state-encoded URLs).
  • Long-running investigation improvements. Requests that exceed the execution time limit now route to a summarizer instead of stalling, so you should receive a response even when an investigation runs long.
  • Fixed account resolution from certain alert URLs, including HURL state-encoded URLs, that previously prevented Autopilot from determining the correct account.
  • More accurate answers, with fewer wrong-account and wrong-entity mistakes: fixes corrected cases where a question about one account or entity could silently return results for a different one, where broad system questions only looked at the literally named service, and where health-summary responses cited correlations that weren't actually supported by evidence.
  • Fewer broken or garbled responses: fixes addressed responses rendering as raw text instead of formatted output, broken charts from bad account data, incorrect classification of unhealthy entities as healthy, and bugs that broke generated queries or crashed the panel.
  • Faster, more resilient responses under load: this release added retry logic for rate-limit errors, a faster model and timeout handling for simple queries, and graceful degradation so an investigation still returns results if an agent is unavailable.
  • Accessibility fixes to composer tooltips and controls in the chat panel.