Lenses and metrics

This page helps you read an Insights number correctly. For every lens it lists the headline numbers and the panels below them, what each one counts, and the line its ⓘ shows in the app.

How to open Insights, filter it and drill into a number is on Insights.

How Insights counts

A few rules hold on every lens, so they’re here once:

• Times are medians, never averages. One ticket stuck for a month would drag an average far from what’s typical. A time also gives the slow end, as “Most within …”: 85 tickets in 100 took no longer than that.
• Shipped means the ticket reached a column whose role is completed, such as Done, inside the window.
• A person is someone acting as a member. By Dawn’s own reviewer answering a mission checkpoint never counts as a person.
• A ticket was rescued (it needed help) if, after an agent run started, a person sent a message while a run was going, moved it back to Todo or Building, or started a new run after one failed. Autonomous means it wasn’t.
• Est. cost is what the tokens would cost at the API’s list prices. Claude seats are billed at a flat rate, so it’s never your bill.
• No data isn’t zero. A number By Dawn has nothing to count for yet shows “No data yet” and a dash.

Overview

Overview and Flow are on every plan.

Number	ⓘ in the app	What it counts
Shipped	“Tickets finished in this period.”	Tickets that reached a completed column in the window.
Cycle time	“Time from starting work to first review, per shipped ticket.”	From the first move into Building (or the first agent run, if earlier) to the first move into Review or Done. A mission checkpoint or a tracker sync moving the card doesn’t count. It’s the same as the ticket’s own total on its page.
Autonomous ship rate	“Share of shipped tickets that needed no human help.”	Shipped tickets that weren’t rescued.
Run success	“Share of runs that completed. Cancelled or stopped runs aren’t counted.”	Of the agent runs started in the window, the share that completed rather than failed.
First-pass rate	“Share of shipped tickets done in one run, with no fixes.”	Shipped tickets with no new run after they first reached Review, and no failed CI after that.
Est. cost / ticket	“Estimated cost of all runs, per shipped ticket.”	Everything the tickets shipped in the window ever cost, divided by how many there were.
Human response	“Time for a person to answer an agent’s question.”	From an agent asking a question, a permission or a checkpoint decision to a person answering it.
Human touches / ticket	“Messages, card moves and decisions by people, per shipped ticket.”	Messages, card moves and checkpoint decisions by people, per shipped ticket.
Questions asked	“Questions asked in Questions, follow-ups included.”	See Questions.

The panels:

• Shipped per day (or week, or month): shipped tickets, split into Autonomous and Human-assisted.
• Where the time goes: the time shipped tickets spent in Todo, Building and Review, split by who held the work: Agent working, Waiting on a person, Checks running or Idle. When two apply, the first in that order wins, and time in Review that nobody else holds counts as waiting on a person. A line under it points at the biggest wait, such as “People held the work 40% of the in-flight time”, with a link to those tickets.
• Cycle time: one dot per shipped ticket, by when it shipped and how long it took.
• Agent activity: when agents run, by weekday and hour, in run-hours, in your organization’s time zone. Runs after a ticket first reached Review aren’t counted.
• Models: shipped work by model, with runs, run success and cost per shipped ticket. A ticket counts for the model of its last build run.
• Needs attention: open tickets that have sat in their stage longer than most of the project’s tickets take from start to review (its 85th percentile over the last 90 days). vs usual says by how much.

Flow

Number	ⓘ in the app	What it counts
Shipped	“Tickets finished in this period.”	As on Overview.
Cycle time	“Time from starting work to first review, per shipped ticket.”	As on Overview.
Lead time	“Time from ticket created to shipped.”	From the ticket being created to it reaching a completed column.
Work in progress	“Tickets in Todo, Building or Review at the end of each day.”	Tickets in those stages at the end of the window, or now.
Reopen rate	“Share of shipped tickets reopened within 14 days.”	Tickets that reached a completed column in the window and left it again within 14 days.

The panels:

• Cumulative flow: how many tickets were in Todo, Building, Review and Completed over the window. Completed counts from the start of the window. On a weekly or monthly view, each point is the end of its week or month.
• Cycle-time distribution: shipped tickets by cycle time, with markers for Typical (the median) and Slow (the 85th percentile).
• Aging work in progress: one dot per open ticket, by stage and how long it has been there, marked On track or Running late on the same rule as Overview’s Needs attention.
• By project on All projects, or This project: shipped tickets and cycle time per project, with links to each project’s cycles.

Agents

Agents, People, Cost and Questions need a plan with analytics dashboards.

Number	ⓘ in the app	What it counts
Runs	“Agent runs started, grouped by how they ended.”	Agent runs started in the window.
Run success	“Share of runs that completed. Cancelled or stopped runs aren’t counted.”	As on Overview.
First-pass	“Share of shipped tickets done in one run, with no fixes.”	Overview’s First-pass rate.
Rescue rate	“Share of shipped tickets that needed a person to step in.”	Shipped tickets that were rescued.
Waiting on a human	“Run time spent waiting on a person’s answer or permission.”	Time runs spent waiting on a question, a permission or something outside By Dawn. Waiting for a usage limit doesn’t count.
CI green first run	“Share of commits whose first finished CI run was green.”	For each pull request commit, whether its first finished CI run passed.
CI reruns	“CI runs restarted at the same commit, by SAW, the worker or a person.”	Split into By Dawn’s CI gate, the worker and a person.

The panels:

• Runs per day, by outcome: runs started, by how they ended: Completed, Failed / error, Stalled, Cancelled by a human, Stopped between turns or Superseded by a newer session. Runs still going appear in Runs but not here yet.

• Is finish confidence honest?: how often a finish the agent rated Green, Yellow or Red shipped on the first try. A Green that often doesn’t is worth a look.

• Mission steps & retries: every mission step that ran, most failures first:

  Column	ⓘ in the app
  First try	“Share of tickets where this step passed on its first try.”
  Tries to pass	“How many tries a ticket needed before this step passed.”
  Failures	“Tries at this step that didn’t pass.”
  Time to pass	“Time from a step’s first try to its pass, when it needed more than one.”
  Pass	“Share of this step’s checks that passed.”
  Flaky	“Share of this step’s checks marked flaky.” A check that passed at a commit where it had already failed.
  Overridden	“Share of this step’s checks a person overrode.”

  Open a row to see why the step was tried again, and click a reason for its tickets.

• Retries by reason: why steps were tried again, over time, such as Check failed, Conflict with main, Branch out of date, Sent back for red CI or Sent back by a person.

• Machines & seats: how busy each machine is (Room, Busy or Full, by the share of its job slots in use) and how close each Claude seat is to its usage limit (OK, High or Exhausted). Busy and High start at 75%. Remote agents → opens the machines.

• Seat headroom: usage of the three busiest seats over the window, against the line where a seat is exhausted.

• Most-used skills & tools: the eight By Dawn skills, subagents, commands and tools used most.

People

People shows where people unblock agents. It’s never a leaderboard: everyone sees their own numbers, and the team table is listed by name, never ranked. Who can see other members is on Team numbers in People.

Number	ⓘ in the app
Reviews	“Tickets this person moved out of Review.”
Review latency p50	“How long agent work waited in Review for this person.” It runs from the ticket’s latest move into Review to this person moving it out.
Decisions	“Mission checkpoints this person decided. Answers SAW’s reviewer gave are not counted.”
Decision time p50	“Time a person took to decide (not SAW’s reviewer). Under 5 s: possible rubber stamp.”
Steers	“Messages this person sent to a running agent.”
Runs dispatched	“Agent runs this person started.”
Questions asked	“Questions this person asked in Questions, follow-ups included.”

The panels:

• My review latency: your typical wait in Review over time (You), next to the organization’s (Org).
• Where I unblock agents: your reviews, decisions, steers and runs started.
• When some of your decisions took under five seconds, a line says how many, with See them: “Fast is fine — just make sure the plan was read.”
• Team: the same numbers for each member, when you can see them.

Cost

Every dollar figure on this lens is an estimate at API list prices, shown with a ~. Tokens from a model with no list price aren’t priced, and the lens says how many there were.

Number	ⓘ in the app
Est. cost	“Estimated at API list prices. Seats are billed flat, so this isn’t your bill.”
Tokens	“Tokens used by agent runs.”
Cache share of input	“Share of input tokens read from cache.” Cache reads cost about a tenth of fresh input, so higher is cheaper.
Est. cost per shipped ticket	“Estimated cost of all runs, per shipped ticket.”
Questions cost	“Estimated cost of answering questions, as reported by the answering agent.” It isn’t part of Est. cost.

The panels break the cost down by model, by project, by epic (the epic the ticket was in when the tokens were spent), by member (who started the runs, listed by name), and by Claude account (the account each run was claimed on). Without permission to see other members, By member shows only You, Other members and Unattributed. Most expensive tickets lists shipped tickets with what they cost against the typical one, and Token mix splits the tokens into cache reads, cache writes, input and output.

Questions

This lens counts how people use Questions, never what they asked: By Dawn doesn’t read the text for it. The epic, cycle, label and runtime filters don’t apply here. Member is the person who asked.

Number	ⓘ in the app	What it counts
Questions asked	“Questions asked in Questions, follow-ups included.”	Every question, first or follow-up.
Conversations	“New conversations started in Questions.”	Including ones since archived.
People asking	“How many different people asked a question.”
Follow-up rate	“Share of conversations where someone asked a follow-up.”
Answer rate	“Share of finished questions that got a full answer.”	A question still unanswered an hour later counts as abandoned.
Answer time	“How long an answer took to arrive.”	Answered questions only.
Became a ticket	“Share of conversations turned into a ticket.”
Questions cost	“Estimated cost of answering questions, as reported by the answering agent.”
Questions tokens	“Tokens spent answering questions, not counting cache.”

The panels show questions asked over time (First question and Follow-up), How questions ended (Answered, Stopped, Failed, Abandoned or In progress), and the same numbers By model, By project and By answer level (Plain, Balanced or Technical).

FAQ

Why do Cycle time on Insights and the ticket's own time agree?

They’re the same measure on purpose: from starting work to first review. So a ticket that looks slow on its page is slow on Insights too.

Why doesn't Cycle time run to Done?

It stops at the first review because that’s when the agent’s work is ready. The wait in Review is measured separately, under Where the time goes and in Review latency on People, so it doesn’t hide inside one number. Lead time runs to Done.

Some numbers on my lens aren't on this page, or the other way round. Why?

A few numbers, such as Human response, the CI numbers on Agents and Questions cost, need a recent version of the Insights service. When it doesn’t send them, the lens leaves them out rather than showing zero.