DORA incident precedent search + classification agent. Gemini 3 + Google Cloud Agent Builder + Elastic MCP. Multi-step mission: search similar past incidents via Elasticsearch, classify using DORA Art.18/19 thresholds, explain via Gemini, build defensibility record, propose storage (gated on approval).
IncidentIQ defines 18 tools with explicit schemas and descriptions. Most tools have clear, domain-specific names (search, classify, analyze, executeProposed) and detailed descriptions (100-300 chars). Input schemas are present and typed. However, several critical gaps reduce the score: (1) Output schemas are NOT documented, tools return objects but the response structure is not specified, forcing LLMs to infer field names; (2) Error handling is minimal, no recovery guidance, no categorization of retryable vs fatal errors; (3) Parameter descriptions lack constraints (enums, ranges, formats), e.g., 'incident' param accepts an object but valid fields are not enumerated; (4) No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear read/write semantics; (5) Composition issues, analyze and analyzeFromDetection overlap significantly, and executeProposed requires a 'payload' from analyze but the payload structure is not documented. Tools like queryAgentEngine and explainViaAgentEngine are vague about what 'full text response' means.
Lightweight reachability probe for Vertex AI Agent Engine (does not run a full query). Used by /health endpoint.
Steps 1-4: search precedents + classify (with recurrence) + draft submission. Read-only multi-step analysis. Embeds incident description, finds similar historical incidents, applies DORA thresholds, checks recurrence rule, and generates Gemini-based explanation + DNB submission draft.
Cross-app handoff: DynaCompliance (real-time detection) → IncidentIQ (classify + file). Accepts a detection payload and routes through the same analyze pipeline.
Deterministic DORA Art.18 major-incident classification. Applies thresholds (>10% clients affected >2h, >€5M transaction value, data breach, core-banking downtime >2h, payments downtime >30min) and recurring-incident aggregation rule.
Count total incidents indexed in Elasticsearch.
Execute ES|QL queries against Elasticsearch. Used for aggregations and complex queries (e.g., listing open incidents, counting by status).
Output schemas not documented. Tools return objects (e.g., analyze returns classification verdict, precedent signal, drafted submission) but response field names, types, and structure are not specified. LLMs cannot reliably extract or chain results.
Parameter constraints missing. 'incident' object parameters accept arbitrary fields; valid fields (clients_affected_pct, transaction_value_eur, duration_min, etc.) are listed in description but not as schema enums or required/optional arrays. LLMs cannot validate input before calling.
No error handling guidance. Tools lack recovery hints, error categorization, or actionable messages. E.g., if Elasticsearch is unreachable, what should the agent do? Retry? Fallback? No guidance provided.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Step 5: human approved → store classification + obligations + alerts. Writes the classification verdict, reporting obligations, and audit record to Elasticsearch. Gated on human approval.
Explain + draft the DNB submission via the deployed Agent Builder agent. Routes the explain + DNB-draft step through Vertex AI Agent Engine (Gemini 3 + Elastic MCP).
Hybrid kNN + keyword search for similar incidents. Uses Elasticsearch dense_vector (768-dim) for semantic search combined with keyword matching.
Shared obligation ledger (JSON; ?format=csv to export). Returns DORA reporting obligations for a company.
Fetch open incidents for the dashboard via ES|QL. Returns incidents with status='open' or 'analyzing'.
True only when the partner Elastic MCP server is actually connected. Used by /health endpoint.
Lightweight reachability probe for Elasticsearch REST connectivity.
Send a message to the deployed Vertex AI Agent Engine (Google Cloud Agent Builder) and return its full text response. The engine runs Gemini 3 + Elastic MCP toolset.
Store audit/defensibility record in Elasticsearch. Records the classification decision, thresholds triggered, precedent signal, and confidence for regulatory defensibility.
Store reporting obligations in Elasticsearch. Called after human approval to record DORA Art.19 obligations.
Count similar incidents on the same service within a reference window (default 30 days). Used to detect recurring-incident aggregation.
Precedent search through the partner MCP server's search tool. Hybrid kNN + keyword DSL search over Elasticsearch incidents index.
Tool annotations missing. Despite clear read/write semantics (search, classify, analyze are read-only; executeProposed, recordObligations, recordAudit are write), no readOnlyHint or destructiveHint annotations are declared. Agents cannot infer safety properties.
Overlapping tool responsibilities. analyze and analyzeFromDetection both classify incidents; the distinction (one from description, one from detection payload) is unclear. Agents may pick the wrong one. Consider merging with a 'source' parameter or clearer naming (e.g., analyzeIncident vs analyzeDetection).