Software companies

Where we come in

Often it starts with a prototype someone built in an afternoon with an AI tool, and a team that needs it to become a real product.

• Plan the real product

  Work out what the prototype is really for, who uses it and what it has to do well, before anyone writes production code.

  Compass→
• Try the options before you build one

  Lay out the shapes the feature could take, put them in front of the people who would use it, and take the one that holds up into the build.

  Prism→
• Sketch it where everyone can see

  Work the thing out on a shared canvas with the engineers, the designers and whoever owns the outcome, then present that same board when it needs a decision.

  Napkin→
• Build the AI into it

  Design the agents and flows on a canvas, then call them from your own code, or run them inside your own systems with the SDK.

  Workbench→
• Know it works before customers do

  Score every change against the bar it has to clear, in CI on every pull request, so a regression never reaches production.

  Caliper→
• See whether it changed anything

  Write down what you expect the feature to move before it ships, then let the numbers from your own systems settle it once it's live.

  Ledger→

• We turned Builders and Backers' prototype into the product their customers use and their own team runs.

  Their team had a prototype that showed what they wanted. We built the real thing around it — the product their customers use, the tooling their team runs it with, and the agents that carry the multi-step work behind both — to the standards something that has to grow and be maintained needs. As we finished, we helped them find the head of engineering who took it over.

  

What your engineers will ask

• Runs on your own infrastructure

  The Workbench SDK runs flows entirely inside your own process, with your own model keys.

• Pin the version you ship

  Every flow keeps its versions, and your code can pin the exact one it runs.

• Evals in CI

  Caliper scores each pull request against your dataset and rubric, and the build fails when the score drops.

• From your editor

  Connect Claude, Cursor or your own agent to the workspace over MCP and work from there.

Three ways of working with us:

• Inside your team

  We join your team to find the work worth changing, build it with you, and prove it holds before it scales.

• Alongside you

  We work with your team a couple of times a week, hands-on with the technical problems and the strategic ones.

• With our tools

  Your whole team works in the tools we use on our own client work, in one shared workspace.