HiveBoard — Analytics Deep Dive Guide #
Version: 0.2.0 Last updated: 2026-02-15
Where are the patterns, costs, and problems? Six analysis sections, one scrollable page.
Table of Contents #
- Overview
- Toolbar and Controls
- Fleet Status
- Cost Rankings
- Activity Rankings
- Error Analysis
- Prompt Analysis
- Tool & Action Usage
- HiveMind Analysis
- Data Sources
1. Overview #
The Analytics page (analytics.html) is a scrollable, analytical view of your fleet's operational data. Unlike the real-time Fleet dashboard, Analytics focuses on aggregated patterns and comparisons over a configurable time range.
The page is organized into six collapsible sections, each addressing a specific analytical question. Each section can be expanded or collapsed by clicking its header.
API data sources: This page reads from the Insights Engine endpoints (/v1/insights/*), the agents endpoint (/v1/agents), and the timeseries endpoint. Data auto-refreshes every 60 seconds, with a "last updated" ticker in the toolbar.
2. Toolbar and Controls #
The toolbar at the top of the page provides:
| Element | Description |
|---|---|
| Title | "Analytics Deep Dive" with search icon |
| Range selector | Dropdown to choose the analysis window: 1 hour, 6 hours, 24 hours (default), 7 days, 30 days, 90 days |
| Last updated | Shows how many seconds since the last data fetch, updated every 5 seconds |
Changing the range triggers a full data reload of all six sections.
3. Fleet Status #
Section tag: LIVE
Question answered: "Who's alive right now?"
This section gives a real-time snapshot of agent health, grouped by operational state.
3.1 Status Strip #
A horizontal bar segmented by state, showing proportional representation:
| Segment | Color | Meaning |
|---|---|---|
| Running | Green | Agents actively processing tasks |
| Idle | Amber | Agents alive but not working |
| Stopped | Red | Agents that are erroring, stuck, or have exceeded their heartbeat threshold |
3.2 Agent Status Rows #
Each agent gets a detailed row showing:
| Element | Description |
|---|---|
| Heartbeat dot | Animated indicator — running (green pulse), idle (amber), stopped (red) |
| Agent name | The agent_id |
| Status badge | Running/Idle/Stopped with color coding |
| Heartbeat age | Time since last heartbeat with heartbeat SVG icon |
| Last event | Event type and time of the most recent event from this agent |
| Cost | Total LLM cost for this agent in the selected range |
| Cost/task | Computed average cost per task |
Agents are sorted: running first, then idle, then stopped.
3.3 Cost by Status #
Three summary cards showing total cost attributed to each status group:
- Running agents' total cost and percentage of fleet spend
- Idle agents' cost (spend accumulated before going idle)
- Stopped agents' cost
Each card shows: total cost, percentage of fleet total, average cost per agent, and average cost per task.
3.4 HiveMind Commentary #
An automated analysis block summarizing the fleet state: how many agents are running, what percentage of spend they represent, and which agents are stopped.
4. Cost Rankings #
Section tag: A1–A3
Question answered: "Who's spending what?"
This section ranks agents by LLM cost and helps identify cost outliers.
4.1 KPI Cards (4 cards) #
| Card | Shows |
|---|---|
| Most Expensive | Agent with highest LLM cost, with call count and token totals. Ranked #1 |
| Least Expensive | Agent with lowest cost. Useful as a baseline |
| Fleet Average | Average cost across all agents with fleet total |
| Cost Spread | Max/min ratio (e.g., "4.2×") showing how uneven spending is. Also shows max vs. average ratio |
4.2 Cost Distribution Strip #
A horizontal stacked bar showing each agent's share of total fleet cost as a colored segment. Segments are proportional to cost and labeled with agent name and percentage when wide enough.
4.3 Ranked Bars #
Horizontal bar chart ranking agents from most to least expensive. Each bar shows the dollar amount and percentage of fleet total.
4.4 HiveMind Commentary #
Identifies the most expensive agent, compares it to the fleet average and least expensive agent, states its share of total spend, and identifies its primary LLM model.
5. Activity Rankings #
Section tag: A4–A6
Question answered: "Who's doing the most work?"
5.1 KPI Cards (3 cards) #
| Card | Shows |
|---|---|
| Most Active Agent | Agent with the most completed tasks. Shows tasks/hr average and success rate |
| Least Active | Agent with fewest completed tasks |
| Fleet Total | Total tasks across all agents, average per agent, and peak hour |
5.2 Ranked Bars #
Horizontal bar chart ranking agents by tasks completed. Each bar shows count and fleet percentage.
5.3 Drilldown: Top Agent #
Two togglable views for the most active agent:
By Task Type: Bar chart showing distribution of work across task types (e.g., "lead_processing", "email_send"). Includes an insight card highlighting the dominant task type.
By Action: Bar chart showing the most-called actions/tools for this agent. Includes an insight card highlighting the most-called action.
6. Error Analysis #
Section tag: A7–A10
Question answered: "Where are things breaking?"
6.1 KPI Cards (4 cards) #
| Card | Shows |
|---|---|
| Most Errors | Agent with the highest error count, with error rate percentage. Shows "0 errors — All clear!" when no errors exist |
| Fewest Errors | Agent with fewest errors |
| Fleet Error Rate | Total errors / total tasks as a percentage |
| Top Error Type | The most frequent error type (e.g., "TimeoutError"), with occurrence count and percentage |
6.2 Ranked Bars #
Horizontal bars ranking agents by error count. Bar opacity decreases with rank to create visual emphasis on the worst offenders.
6.3 Drilldown: Worst Agent #
When errors exist, a detailed drilldown of the most error-prone agent shows three mini-breakdowns:
| Breakdown | Shows |
|---|---|
| By Error Type | What kinds of errors occur (e.g., TimeoutError, ValueError) |
| By Task Type | Which task types produce the most errors |
| By Tool/Action | Which tools/actions fail most often |
6.4 HiveMind Commentary #
Identifies the worst offender, its share of total errors, the dominant error type, and the specific task type and tool where errors concentrate.
7. Prompt Analysis #
Section tag: A11–A14
Question answered: "Which prompts are biggest, most frequent, and most expensive?"
7.1 Prompt Table #
A ranked table of all LLM call names (prompts), sorted by token size:
| Column | Description |
|---|---|
| # | Rank (with colored rank circle: gold/silver/bronze for top 3) |
| Prompt / Call Name | The name given to the LLM call via task.llm_call() |
| Avg Tokens | Average input token count per call |
| Calls | Total number of times this prompt was invoked |
| Agent(s) | Which agents use this prompt (blue badges) |
| Model | Primary LLM model used for this call |
| Est. Cost | Total estimated cost. Color-coded: green (< $0.50), amber ($0.50–$1), red (> $1) |
7.2 HiveMind Commentary #
Identifies the highest total cost driver (prompt × frequency), and calls out the single largest prompt by token count.
8. Tool & Action Usage #
Section tag: A15–A22
Question answered: "Which tools/actions run most often, and how do they perform?"
8.1 Usage Summary Pills #
A row of compact pills, one per action, each showing:
- Total invocation count
- Action name
- Hourly average rate
8.2 Action Detail Table #
| Column | Description |
|---|---|
| Action | Tool/action name |
| Total | Total number of invocations |
| Used By | Which agents call this action (blue badges) |
| Avg Duration | Average execution time |
| Success Rate | Percentage of successful completions. Color-coded: green (≥ 95%), amber (80-95%), red (< 80%) |
| Peak Hour | Hour of day (UTC) with highest usage |
8.3 Hourly Activity Heatmap #
A grid visualization showing action usage by hour of day (0–23 UTC). Actions are rows, hours are columns, and cell intensity shows invocation volume. The heatmap uses a four-level color scale from light (few) to dark (many).
Hover over a cell to see exact count: "action_name @ 14:00 — 23 starts".
A legend at the bottom shows the color scale from "Less" to "More".
8.4 Weekly Aggregation (7d+ ranges only) #
When the time range is 7 days or longer, an additional table shows action usage broken down by day of the week (Mon–Sun) with totals and a trend indicator (▲ up, ▼ down, — flat) comparing the first and second half of the week.
8.5 HiveMind Commentary #
Identifies the action with the lowest success rate, the slowest action, and the busiest action by invocation count.
9. HiveMind Analysis #
Every section includes a "HiveMind Analysis" commentary block — an automated narrative that synthesizes the data into actionable insights. These blocks:
- Identify the most significant finding in each section
- Provide comparative context (e.g., "3.2× more expensive than the fleet average")
- Call out specific agents, models, or actions by name
- Highlight concerning patterns (e.g., "concentrated in the crm_search tool")
These are computed client-side from the API data and update when the range changes.
10. Data Sources #
The Analytics page pulls from six API endpoints:
| Endpoint | What it provides |
|---|---|
GET /v1/agents |
Agent list with status, heartbeat, metadata |
GET /v1/insights/agents |
Per-agent cost, task counts, model usage, action breakdown, task type distribution |
GET /v1/insights/errors |
Error counts by agent, error type, task type, and action |
GET /v1/insights/prompts |
LLM call names ranked by token size, cost, and frequency |
GET /v1/insights/actions |
Action/tool performance: invocations, duration, success rate, hourly distribution |
GET /v1/insights/timeseries |
Time-bucketed metrics for trend analysis |
All endpoints accept a range parameter that matches the toolbar selector. Data refreshes every 60 seconds automatically.