Agent-observability defines 6 tools with explicit schemas and descriptions. However, critical gaps limit production readiness. Most tool descriptions are adequate (100-150 chars), but parameter descriptions are sparse or missing entirely. Input schemas are present but lack depth: many parameters have no type constraints, enums, or validation guidance. Output schemas are completely undocumented, LLMs cannot predict what fields to expect from responses. Error handling is minimal; no recovery guidance or actionable error messages. The tools follow a session-tracking pattern but lack composition clarity and idempotency guarantees. Naming is clear (verb_noun style), but parameter naming inconsistencies (e.g., 'description' vs 'taskDescription') and missing type suffixes reduce clarity.

• Document output schemas for all 6 tools. Specify field names, types, and nesting. Example: start_session returns {id: string, createdAt: ISO8601, status: 'active'}. This enables LLMs to chain tools and extract required IDs.
• Add parameter descriptions to all inputs. For 'status' enum, explain each value: 'success = tool executed without error; error = tool raised an exception'. For 'durationMs', add constraint: 'milliseconds elapsed, must be >= 0'.
• Add error handling guidance. Document failure modes: 'sessionId not found → call start_session first', 'invalid status → must be one of: success, error'. Return structured error objects with code, message, and recovery hint.
• Declare idempotency and retry semantics. Mark 'log_tool_call' and 'end_session' as idempotent (same input → same result) or add deduplication keys (e.g., sessionId + timestamp) to prevent duplicate logs on retry.
• Add tool annotations (readOnlyHint, destructiveHint) to distinguish read-only tools (get_last_session, check_session, get_session_stats) from write tools (start_session, log_tool_call, end_session). This helps agents plan safely.