Evals for Resolve AI

Rotation participation and paging handoff

Agent attaches to the correct rotation and team, triages before escalation, and hands a human a findings-first page rather than a raw alert.

Signal correlation and severity assessment

Correlates signals across the connected observability stack and assigns severity consistent with the evidence gathered.

Blast radius identification

Names affected services, clusters, and orgs with the evidence that establishes scope.

Noise suppression and routing

Silences low-value or self-clearing alerts, marks auto-resolved cases, and routes the remainder to the right team.

Parallel domain-specialized investigation

Distinct roles (triager, investigator, verifier, mitigator) work concurrently and their threads reconcile into one account.

Theory generation and root-cause selection

Multiple theories are surfaced and distinguished as root cause versus contributing factor rather than collapsed into one guess.

Evidence trail and verification

Each claim carries traceable evidence (logs, traces, commits, PRs) and unsupported claims are not promoted to root cause.

Incident narrative construction

Produces what happened, root cause, key evidence, and recommendations with accurate timestamps and resource identifiers.

Interrogating findings in Workbench

Any finding, theory, or evidence item can be questioned interactively and the agent answers from the investigation record.

Mitigation proposal and approval gating

Production-changing actions are presented for engineer approval with rationale and expected effect before execution.

Redirecting an in-flight investigation

Engineer-supplied context or a competing hypothesis changes the agent's line of inquiry instead of being ignored.

Suggested fixes grounded in the codebase

Recommended changes cite concrete files, lines, or configuration locations in the connected repositories.

Schedule and trigger configuration

Agents fire on a cadence or on production events such as deploys and alerts, and only within their configured window.

Task and skill definition with scope

A skill's declared read sources and post destinations bound what the agent may access and where it may write.

Templates and reuse

Starting from a template (deploy health, daily digest, alert triager, resource report) yields a working, editable agent.

Priority feed and digest output

Recurring output is summarized and prioritized rather than replayed in full.

Connector coverage across the stack

Code, infrastructure, telemetry, knowledge, and collaboration sources are reachable and their absence is stated, not filled in.

Cross-source correlation

Findings join evidence from more than one system (for example a commit or PR against traces and logs).

Tribal knowledge capture

System-specific conventions from connected knowledge sources are applied to investigations rather than generic guidance.

Degraded and missing-source behavior

When an integration is unavailable or returns errors, the agent reports reduced coverage instead of asserting unverified conclusions.

MCP tool invocation from external agents

Telemetry queries and investigation retrieval return structured, referenceable results to a calling agent.

Embedding in existing workflows

The same capabilities are reachable from Slack, CLI, IDE, or another agent with consistent results.

Investigation retrieval and linking

A returned investigation carries its status, evidence count, and a link back to the record in Resolve.

Custom skill authoring

Team-authored skills compose Resolve primitives without reimplementing context, investigation, or remediation.