Kelright

Know whether your engineering team is working well, and on the right things.

The Kelright Benchmark reads your delivery from the systems you already run, against a baseline frozen once. The numbers land in your own Datadog.

Benchmark your team

Opens an email to hello@kelright.com.

Back a portfolio? Read it as one cohort.

Deployment frequency, before and after

Illustrative

Fig. 1Ninety days read back and frozen, then every week after read under the same rules. The line is made up, so it has no scale.

What it reads

Six families of measures, read from the systems you already run.

GitHub for pull requests and deployments, Linear for the work, and your own Datadog for what it already measures, agent spend included. Read-only, with your consent.

What the Benchmark reads

  • Delivery

    How often you ship, and how safely: the four DORA measures.

    • Deployment frequency
    • Lead time for changes
    • Change failure rate
    • Time to restore
  • Review flow

    Where a pull request waits between being opened and being merged.

    • Pull request cycle time
    • Time to first review
    • Review time
    • Merge time
    • Pull request size
    • Pull requests merged
    • Review rounds
  • Work flow

    Whether the work you start is finished, and stays finished.

    • Issue cycle time
    • Issue lead time
    • Time in each state
    • Throughput
    • Work in progress, and its age
    • Cycle completion rate
    • Issue rework rate
  • Agent adoption

    How much of the work agents do.

    Exists once agents write code

    • Agent-assisted pull request share
    • Agent-assisted commit share
    • Issues delegated to agents
    • Merges outside working hours
  • Agent economics

    What agent work costs, and whether it comes back right.

    Exists once agents write code

    • AI spend per merged pull request
    • AI spend per delivered issue
    • Tokens per thousand lines changed
    • AI spend per active engineer day
    • Cycle time, agent-assisted and not
    • Rework, agent-assisted and not
  • Quality

    Whether what merges holds.

    • CI pass rate
    • Tests on agent-written pull requests
    • Rework rate
    • Revert rate
Fig. 2Every measure, by family. The two marked families exist only once agents write code. Each value names its system and rule version.

Before and after

A before that is frozen once, so an after means something.

Better than before is only a claim if the before cannot move.

Read back ninety days
Ninety days of history from each system when you connect, so there is a before on day one.
Frozen once
Frozen once every system is read, and nobody can recompute it, us included. Freeze it sooner and it says which systems were still reading.
One set of rules
Every value is computed under a named rule version, and the after always under the before's.
When the rules change
The original window is re-measured as a second before beside the first, which is never rewritten.

The right things

Working well is speed. Working on the right things is finishing what you set out to do.

Fast delivery of the wrong work still shows in the numbers, if you read the right ones.

Planned work, finished
The share of a cycle's planned work done by its end. Fast and never finishing is busy, not well directed.
Work that stays done
Issues reopened and changes reverted. Work that comes back was not right the first time.
Where work waits
Time in each state and the age of work in progress: the review queue, the ticket nobody owns.
Spend that lands
Agent cost per piece of work that reaches done, not per line written.

Your company decides what the right thing is. The Brain keeps those decisions and holds agent work to them. The Benchmark reads the work with it and without, so what the Brain saved is a number here, not a claim there.

Before it starts

A planned project costed from what your own work has cost, and scored when it closes.

Once it knows what each size of work costs your team, agents and reviewers together, the Benchmark reads a planned project forward as a range. A forecast nobody scores is a guess with a date, so every one is kept and scored.

A range, from your own history
Your project's issues, costed from what each size has cost you: a twentieth-to-eightieth percentile range, assumptions beside it. No price list, no other team's data.
Scored when it closes
Kept with its date and compared with the final cost, so you can read how well it has predicted before you lean on it.
What a budget buys
The other way round: how many planned issues a budget covers, in dependency order.
With the Brain, and without
The same work costed both ways: what the Brain saved, net of its own cost, with the sample size beside it.

Against peers

See where your team sits beside teams that already work with agents.

Every team is read under the same rules. Peers are combined, none is named, and your numbers go only where you choose.

Where would your team sit?

Illustrative

Fig. 3Each dot is a team that works with agents, and every position is made up. The better end of each line is on the right.

Across a portfolio

One set of rules across a portfolio, combined by default.

A fund reads its cohort as one, quarter after quarter: combined, no company named, no person in the data. A company’s own numbers reach the fund only if it opts in, in writing.

A cohort, from its baseline to now

Illustrative

Fig. 4A fund's cohort, baseline to now: the middle half and the median, combined, no company named. Every position is made up.
Bring the Benchmark to your portfolio

Opens an email to hello@kelright.com.

In your Datadog

It lands in your own Datadog, and what your agents emit never passes through us.

You connect it
From the Kelright tile in your Datadog, then GitHub and Linear, read-only.
Yours to keep
Metrics, a dashboard and baseline markers in your Datadog, under a key you can revoke.
Never through us
Agent spend is read from your Datadog, where it already is. Nothing an agent emits passes through us.
Stop when you like
Disconnect a system and its reading stops; the frozen before stays put.