Measure the ROI of agentic coding

AI coding is here to stay. You need visibility into its impact, cost, and quality to measure ROI — and to continuously improve the agents and workflows that build your codebase.

Complete visibility on every pull request

Accurate, line-level AI-attribution on every commit — the % AI, token cost, and agent sessions behind every pull request.

Track AI's impact on every PR
Pull requestsLast 30 days · main
Pull request% AICostAuthorAgentModel
Rewrite rewrite-ops attribution: object-based core, trace2 ingestion, fuzzer#1542
88% AI
$42.10
SVsvarlamov
Codex
GPT-5.5
Add idempotency keys to webhook retries#1540
81% AI
$24.60
LTLois Tam
Claude Code
Opus 4.8
Fix auth token refresh race condition#1539
74% AI
$13.05
MCMarcus Chen
Cursor
Sonnet 4.6
Migrate billing cron to durable queue#1536
90% AI
$48.20
SVsvarlamov
Codex
GPT-5.5
Add Codex hooks.json install option#1534
71% AI
$11.40
LTLois Tam
Claude Code
Opus 4.8

Track the full lifecycle

See how much AI code actually makes it to production — and the impact it has when it gets there.

AI Code Journey
7.8kdiscarded by agent1.1krewritten in review0.7kPR closed without merge1.4kchurned within 30 days12kGENERATED4.2kCOMMITTED3.1kREVIEWED2.4kMERGED1.0kDURABLE 30d
12 lines generated for every 1 to production
65% of AI-code is discarded before commit
58% of merged code is churned within 30 days
Generated : Production
12:1

lines committed for every 1 still in production

very high
Rework rate
26%

of reviewed lines rewritten before merge

higher than avg
Code churn
58%

of merged AI-code churned within 30 days

high
Avg. Tokens / PR
1.2M

tokens across all agents per merged PR

Tie incidents to agent sessions
Incident
INC-204 · Checkout 500s
spike traced to a bad retry path
git ai blame
Blamed lines
payments/charge.ts
L88–96 · AI-attributed
session
Agent session
Cursor · Opus 4.8
“add retry to charge()”

Track agent autonomy

Straighten the path from prompt to production: fewer human turns, fewer tokens, and agents that have the context they need to finish the job on their own.

promptproduction
31−71%
Human turns
840K−38%
Tokens
211−76%
Tool calls
Autonomy

A straighter path

Fewer human turns, less rework, and fewer regressions between prompt and production.

Fewer tokens

Less wandering and backtracking means the same work ships for a fraction of the spend.

The right context

Agents reach the skills, rules, and docs they need to finish the job on their own.

Token efficiency

How much AI code gets generated for every line that ships — measured per repo and per team, so you can pick your targets and improve agent readiness where it counts.

Generated : Production
30: 1

AI writes ~30 lines for every line that ships

A measure of how much work it takes agents to get code into production.

30 generated1 shipped
Compare across repos & teams
  • payments-api
    Autonomy88%
    Gen : Prod14:1
  • web-dashboard
    Autonomy71%
    Gen : Prod23:1
  • platform-core
    Autonomy58%
    Gen : Prod31:1
  • legacy-billing
    Autonomy32%
    Gen : Prod58:1

Direct your investments at the repos and teams where agents run with the least autonomy and the worst generated:production ratios.

Harness engineering pays for itself

Teams that improve agent autonomy and readiness cut token spend by 30–55% — the same output for far fewer generated lines.

Compare models & workflows

Agent autonomy, token efficiency, and the average cost of a merged PR — side by side across every model and workflow your team uses.

modelAgent autonomyToken efficiencyAvg cost / merged PR
Opus 4.8
94
13:1
$38.50
GPT-5.5
79
27:1
$21.40
Sonnet 4.6
72
17:1
$9.20
Haiku 4.5
43
31:1
$3.40
Also break down by team · repo · contributor

Measure ROI

Get visibility into everything coding agents do, and the impact of the code they contribute.