Where it goes

agents cost answers what you spent. agents insights answers how you worked, and is the only surface that splits any of it by the account that did the work.

By account

The balanced strategy is the default, so runs are already sprayed across every signed-in account. A report that reads one account's directory describes a fraction of the work and attributes all of it to one org — which is why agents insights exists as a separate surface from a single harness's own built-in report.

agents insights                                # last 30 days, by account
agents insights --by project --since 90d       # which repo is eating the time
agents insights --account "Turing Labs" --all  # one account, every session ever
agents insights --json                         # stable contract for a dashboard
agents insights --narrative                    # add a written read on the numbers

What's in the report

• By account — sessions, cost, duration per account (or per --by dimension).
• Top tools / languages / models — histograms by name, read from transcript content.
• Friction — interruptions, tool errors, gap until the next message, and agent silent stalls: the assistant ends a turn while work remains, sits idle, and only resumes when you type "continue" later.
• What changed — lines touched, files created or modified, commits and pushes seen in shell commands.
• When you work — a 24-slot local-time histogram of your messages.
• Parallel sessions — overlapping runs, and how many straddled two accounts.

Parsing every transcript is expensive, so each session's facets are cached and only re-read when the file changes. --refresh discards the cache. --narrative is the only path that reaches a model — it pipes the aggregate, never a raw transcript, through a headless claude -p.

Not this

Three adjacent surfaces answer three different questions, and none of them is a rename of another:

• agents perf — latency. p50/p99 for hooks, commands, and runs. Not popularity, not spend.
• agents usage — live quota headroom right now, per provider. Not history.
• agents sessions stats — which skills and slash commands got invoked. Not behaviour.

agents trends still runs, as a deprecated alias for the counter-mix tree under agents insights mix — harness/model/token distributions without the transcript parse. Reach for the bare command above when the question is about time and cost by account.

Related: Track spend, Cap spend, Sessions.