← Back to overview

For a support team · agent building

What if the plain-English procedure were the software?

The procedure is the source code — what the team writes is what the agent does.

Today building an agent means prompt engineering or wiring flow-charts and tools — the plain-language procedure and the software live apart.

Why now LLMs finally let the person who knows the procedure build the agent — without waiting on engineering.

Use cases Customer support, ops & back-office, internal tools, compliance.

Interactive demo ↗

The question

The people who know the policy write it in prose. Then engineers translate it into flows — and the two quietly drift apart.

The approach

Skip the translation. The procedure, written in ordinary sentences, compiles straight into the agent — sentence by sentence, into typed steps you can see.

The feature

Each sentence becomes a node: ask, verify, condition, action, escalation, never-rule. Edit the prose and the graph updates; the words and the behaviour can’t drift.

Function

The compiled graph

prose-to-agent — compiled graph
Loading…

A conditional written in prose compiles into a branch you can read — condition, yes/no, action.

Function

One sentence, one node

“If the refund exceeds €75, escalate to the returns supervisor with a short summary.”
↓ compiles to
◇ condition↗ escalation

Every sentence maps to one typed step — hover the prose to see its node, hover a node to see its sentence.

Newsletter

Notes from building at the human–AI interface.

Every two weeks: new builds, notes, explorations — the interaction patterns that are working for agentic AI, and short field notes you can put to use.

Subscribe