Your agents call models, invoke tools, touch data, burn tokens and fail — each inside a different vendor’s console. Oliver lands all of that activity in one table at full resolution, so reliability, efficiency, cost and governance are one query away instead of three exports and a spreadsheet.
Claude reports on Claude. Gemini reports on Gemini. Codex reports on Codex. Nothing joins them, so the questions that actually matter — which agents are unreliable, where the token spend went, what an agent touched before it got something wrong — cannot be asked in one place. Meanwhile the coding and enterprise surfaces emit telemetry in different shapes, and the richest of it expires out of a retention window measured in weeks.
Every model call, tool invocation, retry and error arrives as its own row, joined by session. When an agent goes wrong you can walk the sequence that led there instead of inferring it from a summary.
Token counts and cost land beside the agent, model, session and tool that incurred them, so “what did this agent cost us last month” and “which tool is burning the context window” are ordinary queries.
Percentiles by agent, model and tool are pre‑built, so a slow tool or a model that retries twice as often shows up as a number rather than a hunch.
Activity sits next to the audit trail from Agent Flight Control, so what an agent was allowed to touch and what it actually queried can be reviewed together, after the fact.