Control

Analytics

The Analytics dashboard gives you a rolled-up view of your agents' performance across three lenses: overall health, cost, and quality. Use it to spot trends, compare agents, and decide where to focus optimisation effort.

Overview tab

The Overview tab answers is the system healthy overall?

Total Sessions

All sessions in the selected range, regardless of status.

Total Cost

Cumulative LLM spend across all sessions and agents.

Org Error Rate

Sessions with error_count > 0 as a % of total sessions.

Session trend chart

A time-series bar chart showing daily session volume for the selected date range. Each bar is split by status: done (green), stale (amber). Hover a bar to see exact counts. A sudden drop in volume often indicates a deployment failure or a misconfigured API key.

Agent breakdown

Below the trend chart, a per-agent table ranks agents by session count (descending). Columns show: sessions, total cost, error rate, avg steps, and last active timestamp. Use this to identify which agents are driving the most spend or have the highest error rate.

Cost tab

The Cost tab answers where is the money going?

Total Spend

All LLM costs in the range.

Avg Cost/Session

Mean cost per session — useful for budgeting.

Top Agent Spend

The single agent responsible for the most spend.

Cost by agent

A horizontal bar chart ranked by total spend. Each bar shows the absolute cost and the percentage of total organisational spend. Use this to find the agents with the largest cost footprint and prioritise them for prompt optimisation or model downgrade experiments.

Cost by model

Breaks down spend by LLM model (GPT-4o, Gemini Flash, Claude Sonnet, etc.). Useful when your agents use multiple providers: see which model contributes the most to your bill and whether cheaper alternatives are already in use elsewhere.

Top sessions by cost

The ten most expensive individual sessions in the selected range, with direct links to the session drawer. Long-tail outliers — a session that cost $5 when the median is $0.02 — usually indicate a runaway loop or a missing step limit policy.

Tip:If a single session consistently tops the cost list, open it and check the step count in the session drawer. A missing step_limit policy is almost always the root cause.

Quality tab

The Quality tab answers are the agents working correctly?

Org Error Rate

Same as Overview — sessions with errors / total.

Avg Latency

Mean session duration across all agents.

Goal Completion Rate

Sessions with goal_completed=true / sessions with metrics.

Error rate trend

Daily error rate (errors / total sessions) for the selected range. Correlate spikes with deployments, policy changes, or provider outages using the trend chart.

Latency trend

Average agent latency per day, measured from session start to session close. Latency is derived from event.data.meta.latency (seconds) — the duration_ms field is not used and is always 0.

Agent quality scores

A per-agent table showing goal-completion rate and average score (from report_metrics()) for agents that have reported outcome data. Agents without metrics are listed separately. Use this view to identify which agents are underperforming on your defined quality criteria.

Date range and filters

All three tabs share the same date range picker. Presets: last 7 d (default), 30 d, 90 d, or a custom range. Changing the range re-fetches all charts and tables at once.

Note:Analytics data is aggregated server-side and cached. The Overview and Cost tabs update within seconds of new sessions completing. The Quality tab updates when sessions with report_metrics() data close.

Agent health data

Clicking an agent in any Analytics breakdown opens the Agent Health view: daily session buckets, KPI summary (session count, error rate, avg cost), a by-variant breakdown if the agent has run experiments, and the signals log.

When to use Analytics vs Sessions

  • Analytics — understanding trends, comparing agents, budget planning, quality regression detection. Aggregated view, no per-session detail.
  • Sessions — debugging a specific run, investigating a signal, reviewing the event timeline. Per-session detail with full event trace.

Next steps

  • Sessions — drill into individual runs from any Analytics table.
  • Experiments — run a controlled A/B test when Analytics reveals a quality or cost problem.
  • Policies — set cost and step limits based on what you find in the Cost tab.