Defender Impact
Amazon Quick’s agentic incident triage assistant compresses the evidence-gathering, synthesis, and handoff stages of incident response into a single automated workflow — closing the gap between alert firing and structured, documented triage that has historically depended on manual coordination across disconnected tools.
Capability Overview
Amazon Quick has shipped a reference architecture for an agentic incident triage assistant that wires together New Relic’s observability platform (via MCP Server) and Asana through native connectors. From a single engineer prompt, the agent autonomously calls five New Relic reasoning tools: log analysis, alert insights, user impact assessment, transaction analysis, and natural-language NRQL query generation. It then synthesises an RCA brief and creates a tracked Asana task. The entire evidence-gathering and handoff lifecycle is delegated to the agent.
The architecture’s significance for defenders is the consolidation of two planes that are usually manually bridged: read access to live observability telemetry and write access to project management systems, with the LLM as the reasoning layer connecting them. An SRE who previously had to query New Relic, interpret logs, assess user impact, draft a summary, and then file a task — across multiple context switches under time pressure — can now initiate that full sequence from a single conversational prompt.
The natural_language_to_nrql_query tool is particularly notable for platform teams. It lowers the NRQL expertise barrier, allowing a broader range of responders to query production observability data precisely during the moments when speed and accuracy matter most. The multi-connector model means that the context assembled during triage — service topology, error rates, affected transactions — flows directly into the Asana task artifact rather than being summarised imprecisely in a handoff message.
Defensive Advances
This architecture gives defenders several concrete capabilities they previously lacked or had to build manually:
- Systematic telemetry reasoning at scale. The
analyze_entity_logsandanalyze_transactionstools apply consistent analytical reasoning to high-volume log and trace data, surfacing signal that fatigued human reviewers are liable to miss during major incidents. - Consistent incident documentation. Agent-generated RCA briefs tied directly to the telemetry evidence replace ad-hoc summaries, improving the quality and comparability of post-incident reviews.
- Faster, lower-friction observability queries. Natural-language NRQL generation democratises access to production data during incidents, reducing dependence on specialists.
- Automated, auditable handoff artifacts. Asana tasks generated by the agent carry structured RCA content, creating a traceable record of what was known, when, and by what reasoning path.
Residual Gaps
The architecture as described does not yet specify input validation controls for telemetry ingested by the agent. Application logs, alert payloads, and transaction traces are high-volume, partially external-influenced data sources; without a sanitisation layer between raw telemetry and the LLM reasoning step, the agent’s conclusions are only as reliable as its inputs. Teams should treat this as a maturity requirement to address at deployment, not a reason to defer adoption.
NRQL query scope constraints are also not addressed in the reference architecture. Without allowlisting or service-boundary enforcement, queries could return data outside the incident’s intended scope — a coverage gap that matters for organisations with strict data compartmentalisation requirements.
Finally, the multi-connector credential model — a single Amazon Quick Professional-tier session authenticated to both New Relic and Asana — warrants session token protection practices commensurate with the combined access it represents. This is a standard operational maturity question for any multi-platform integration, not unique to this architecture.
Framework Mapping
AML.T0051 (LLM Prompt Injection) and LLM01 (Prompt Injection) are the technique categories this deployment should be hardened against through input validation — the architecture’s telemetry ingestion pipeline is the surface to protect. AML.T0057 (LLM Data Leakage) and LLM06 (Sensitive Information Disclosure) frame the data governance controls needed around RCA brief outputs. LLM08 (Excessive Agency) informs the case for human review gates on agent-generated tasks — a maturity control that keeps autonomous write actions accountable. LLM07 (Insecure Plugin Design) provides the design standard against which the MCP Server integration should be evaluated, particularly around tool input validation.
Deployment Considerations
Teams integrating this assistant into production incident response should consider three operational factors:
Telemetry trust boundaries. The agent reasons over logs and traces that may include externally influenced content. Establishing a validation or filtering layer before telemetry reaches the LLM reasoning step is the primary maturity requirement for production deployment.
NRQL scope governance. Define query scope constraints appropriate to your service topology before enabling the natural_language_to_nrql_query tool broadly. Allowlisting query patterns or enforcing service-boundary filters ensures the agent’s observability access matches its intended incident scope.
RCA brief review cadence. For high-sensitivity production environments, build a human review step into the workflow for agent-generated Asana tasks before they are actioned. This is most valuable for RCA briefs involving external-facing services where the incident narrative may inform customer communications.
Defender Checklist
- Map the assistant to your existing incident response runbooks and identify which New Relic-instrumented services to onboard first
- Configure New Relic connector access as read-only and scoped to relevant service entities at deployment time
- Limit Asana write access to designated incident project boards from day one
- Implement a telemetry input validation layer before logs and traces reach the LLM reasoning step
- Define NRQL query scope constraints or allowlists appropriate to your service boundary requirements
- Establish a human review step for agent-generated RCA briefs in high-sensitivity production workflows
- Enable immutable audit logging for all agent tool calls and outputs to support post-incident forensic review
- Apply session token protection practices to Amazon Quick connector credentials commensurate with their combined observability and task management access