Agent Insights Beta
Visibility into every agent run
Your org bought Claude Code, Cursor, and Codex seats. Adoption dashboards from those tools tell you who logged in. They don't tell you whether a session ended in shipped code or three retries and a dead end. Agent Insights aggregates session-level behavior across every agent your team uses — into one dashboard that shows adoption, spend, and effectiveness, without anyone reading a single transcript.
This is a coaching tool, not a surveillance one — Entelligence never sends raw session transcripts, only computed metrics (token counts, tool-call patterns, retry rates, and similar signals).
🎥 Video demo — coming soon
Adoption isn't the same question as "is it working"
12 of 15 engineers used Claude Code this week.
Those 12 had a 40% retry rate on one project this week — up from 10% last week — which is where the friction actually is.
What's in the dashboard
| Tab | What it shows |
|---|---|
| Overview | Team health at a glance — adoption, activity trends, and a session activity map. Per-member breakdown for drilling into any individual's usage. Coding-behavior signals like tool-call efficiency and retry rates. Skill/command usage across the team. Spend patterns across connected agents. |
| Budget | Project-level budgets versus actual spend, with targets you can set per project. |
| Outcomes | Where agent-driven work actually lands — spend-to-production tracking by outcome category, plus activity themes and heatmaps showing where usage concentrates. |
| Entelligence Wrapped | A personal, shareable year-in-review card per developer — usage patterns, wins, and coding persona — exportable as a standalone page. |
Supported agents
Claude Code, Cursor, Codex, GitHub Copilot, Gemini, Aider, and Windsurf.
How data gets in — and what never leaves the machine
Agent Insights runs on the same engine behind the open-source cinsights project, driven from each developer's machine via entelligence insights in the CLI. It reads session files that already exist locally (~/.claude, ~/.codex) — or connects through Entire.io for cross-machine/cross-agent coverage, or Arize Phoenix for centralized trace observability — scores them locally, and syncs only the computed result up to your org's dashboard.
| Sent to your dashboard | Never leaves the developer's machine |
|---|---|
| Token counts | Raw prompts |
| Tool-call names and counts | Raw agent responses |
| Efficiency and retry scores | Tool input/output (where code, commands, or secrets could appear) |
This isn't a policy promise — the sync mechanism is built around an explicit allow-list of computed fields, so there's nothing to accidentally over-share.
What "efficiency" and "friction" actually measure
Rather than lines of code or session duration — numbers that go up whether a session went well or in circles — Agent Insights tracks behavioral signals that actually correlate with whether an agent session worked:
| Signal | What it catches |
|---|---|
| Tool-call efficiency | Purposeful tool use vs. flailing — how many calls it takes to actually get something done |
| Reads-before-edits | Whether context gets gathered before changes are made, or an agent edits blind |
| Retry rate | How often a task needs multiple attempts before succeeding |
| Context pressure | Early signs a conversation is running up against a model's context limit |
These roll up into per-developer behavior patterns and team-wide health tiles — framed as coaching signals ("here's where the friction is"), not individual performance scores.
Some sections (like detailed outcome-to-production tracking) are still rolling out during the beta — you may see a "coming soon" state on parts of the dashboard depending on your org.
See Setup to get your team connected.