Compare agent-authored work with human work on the same measures.
“Agents are writing code now. Is that work reaching anything that matters?”
Agent pull requests are read beside human ones on share of merged work, cycle time and cost. Each is linked to the goal it served.
Income verified through open banking, not payslips
AI spend $56.25 on 3 pull requests · $2.30 matched by push
Work in progress
Self-employed applicants are the third of the manual income queue that phase one left behind.
- In ProgressStoryMORT-531Detect self-employed income from twelve months of transactionsRahul Shah
- In reviewPull requestmortgage-origination #648Self-employed income from transactionsagent: ClaudeAI $38.40
- OpenPull requestmortgage-origination #651Count regular transfers from a business account as incomeagent: Cursorsuggested, not confirmedAI $11.75
- To DoStoryMORT-532Show the applicant which transactions we counted as income
- DoneBugMORT-533Consent screen times out on slow bank redirectsJoe Barnes
- MergedPull requestmortgage-origination #646Longer timeout on the bank consent redirectJoe BarnesAI $6.10, $2.30 matched by push
The work under one goal, who shipped it — person or agent — and what each run cost.
- People and agents on one boardLive
Work under way by people and by agents appears on one board, grouped by the goal it serves.
- Agent and human pull requests comparedLive
See the share of merged pull requests written by agents, and their cycle time beside the human figure.
- Cost of a pull requestLive
Each pull request shows what its AI spend and people time came to, without timesheets.
- Session spend tied to a branchLive
Each Claude Code session's spend arrives tied to its repository and branch, so cost joins the pull request.
- AI spend by goalLive
See how much AI spend reached each goal, and how much reached none.
- Agent work linked to its goalLive
An agent's pull requests come back linked to the goal and the person who sent the work.
Live In the product today. Preview Working, refined with customers.
Kestrel's CTO is asked whether coding agents are worth their cost.
- 01Agents wrote twenty-two percent of the quarter's merged pull requests on the lending platform, with a one-day median cycle time.
- 02The average agent pull request cost about a tenth of a comparable human one, counting AI spend and review time.
- 03Two thirds of agent work served the time to offer goal. A fifth of AI spend was linked to none.
Kestrel Bank is a fictional example. The figures are calculated by the product.
- GitHub
- Claude Code
- Anthropic
- Cursor
- Cursor agents
- Devin
- Claude
If a system you rely on is not listed, ask for it. Outcome’s integration builder drafts the connection, and it is offered to you to try before it is switched on. How the agents work
The comparison uses your own teams' figures over time.
Leaders can decide where agents are worth extending, using the cost and the result of work already shipped.
More in engineering excellence- Delivery metricsDeployment frequency, lead time and pull request flow, by goal.
- AI adoptionWhich teams use which AI tools, and what it costs.
- Agents at workHand a goal's work to coding agents, with human approval.
- ReliabilitySentry errors and user feedback, grouped and given an owner.
- Where effort goesHow much work reaches a goal, and where it overlaps.