A multi-agent orchestration system for incident triage, root cause analysis, correlation, notification, and human-in-the-loop approval. Implements a 5-agent pipeline: Triage Agent (severity classification), Correlation Agent (runbook + past incidents), RCA Agent (root cause analysis), Notifier Agent (Slack formatting), and HITL Gate (human approval).
12 tools with mixed quality. Most have descriptions (good), but schemas are incomplete or inferred. Tool names follow verb_noun convention (query_, search_, get_, classify_, correlate, draft_, format_, evaluate, orchestrate, process_, log_). However, many parameter descriptions lack detail on constraints, formats, and valid ranges. Output schemas are largely undocumented. Error handling is minimal, no recovery guidance or categorization. The HITL gate (evaluate) and orchestrate tools are complex but lack clear error paths. Telemetry tools (log_agent_trajectory, log_span) are well-intentioned but add noise without structured observability integration.
Classify alert severity and extract key metadata.
Fetch runbook and past incidents for a triaged alert.
Draft an RCA from correlation + triage data. Returns {rca_draft: str}.
Evaluate whether a remediation requires human approval. SEV1/SEV2 -> blocks on input(). SEV3 -> auto-approves (configurable). Returns {human_required: bool, approved: bool, action: str, severity: str}
Format incident data as a Slack notification. Returns {slack_message: str}.
Retrieve a Sentry error event by ID (simulated for demo).
Simulates structured logging and tracing for a production agent. In a real production environment, this would integrate with OpenTelemetry and Google Cloud Trace/Logging.
Output schemas undocumented. Tools return dicts (e.g., correlate returns {runbook_summary, past_incidents, service, severity}) but no formal schema is declared. LLMs cannot plan downstream calls without knowing return structure.
Parameter descriptions lack constraints. 'query' in query_developer_knowledge has no length limit, format, or guidance. 'keywords' array in search_github_issues has no min/max items. 'amount' in process_expense has no range (should be 0 - 999999). LLMs will pass invalid values.
Error handling absent. No tool documents what happens on failure (API timeout, invalid input, permission denied). evaluate() blocks on input() for HITL but has no timeout or fallback. orchestrate() chains 5 agents but no error recovery path if one fails.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | <=2025-11-25 | v2 |
Log a telemetry span. Never raises - telemetry must not crash the agent.
Run the full 5-agent triage pipeline for a single Sentry event.
Process an expense and determine if human approval is required based on amount threshold ($100).
Query Google Developer Knowledge MCP for runbook/docs content.
Search GitHub issues for past incidents matching keywords.
Telemetry tools (log_agent_trajectory, log_span) are observability helpers, not domain tools. They add 2 tools that don't advance incident triage. Consider moving to server-side middleware or OpenTelemetry integration instead of exposing as callable tools.
process_expense tool is out of scope. This is a capstone project for SRE incident triage, yet process_expense (expense approval) is unrelated. Suggests incomplete tool curation or copy-paste from a template.