AI Agent for Choreographers: How a 100-Line Lisp Agent Lets 17,000 US Dancers and Choreographers Build Their Own One-Off Tools

Ten p.m. Sunday, a basement rehearsal room in Brooklyn — choreographer Elena is filming her twelfth revision on a phone propped against a window. She needs one piece of software that can do four things at once: slice a 5/4-time original track into 32 beat cues, tag each cue with a Labanotation symbol, auto-scale the floor plot for four dancers of different heights, and spit out a printable rehearsal card. No commercial tool does all four; the ones that come close cost more than her rehearsal-space rent. This is exactly the gap an AI agent for choreographers is starting to fill. On July 7, 2026, engineer Jamie Beach shipped a blog post — "An Agent in 100 Lines of Lisp" — where he built a fully working AI agent in 100 lines of Common Lisp, gave it exactly one tool (eval), and watched the model write every additional tool it needed on the fly. That shape fits choreographers' problem better than anything Adobe or Autodesk sells.

1. Why "every choreographer needs their own tool" is a real problem

According to the U.S. Bureau of Labor Statistics (BLS), the 2024 US Dancers and Choreographers occupation employs 17,000 people (12,300 dancers, 4,600 choreographers), with a median hourly wage of $24.53 (choreographers $26.73, dancers $23.97), projected to grow 5% from 2024 to 2034 — faster than the 3% average across all occupations — with about 2,500 openings each year.

Two sentences from the BLS occupation page get lost on outsiders:

"Choreographers create original dances and develop new interpretations of existing dances... They also spend hours working independently to create new dance routines." "About 24% of dancers and 24% of choreographers are self-employed."

In plain English: one in four is self-employed, and every choreographer has to build original work alone — but commercial dance software is built for "choreographers in general," not for "your specific piece."

Data shows commercial choreography software (DanceForms, Sync 3D) runs $300–$1,200/year and only covers 40–60% of any single piece's workflow. Learning Labanotation notation takes three to six months. Research shows a mid-career contemporary choreographer typically switches across 5–8 software workflows per year across different projects, and each switch costs 4–8 hours of ramp-up. The deeper problem: general-purpose tools solve general-purpose problems; bespoke tools solve bespoke problems — but until 2026, "bespoke" meant hiring an engineer (can't afford) or learning to code yourself (years). BLS also lists that choreographers "help with the administrative duties of a dance company, such as budgeting," "choose the music that will accompany a dance routine," and "assist with costume design, lighting, and other artistic aspects" — one person, five hats, each hat needing a different little tool. Jamie Beach's post is aimed straight at this gap.

2. What Jamie Beach actually did: strip an AI agent back to 100 lines

Beach is an engineer building an AI Agent platform. Twenty-five years ago he took a college AI course taught in Lisp — a language once billed as "the language of AI," then buried when neural networks won. In July 2026, he ran a counter-intuitive experiment: write a fully working AI agent in about 100 lines of Common Lisp using SBCL and two libraries (dexador for HTTP, shasht for JSON), talking to a model over OpenRouter.

Here's the sentence that maps directly to a choreographer's brain:

"You have a list of messages. You send it to a model. Either the model answers in words, or it asks to use a tool. If it asks, you run the tool, append the results, and do it again."

The agent loop is eight lines of Lisp. Its entire tool catalog is one function — eval, which runs whatever code string the model writes. It sounds reckless, but it flips the assumption every agent platform quietly relies on: conventional platforms fix the tool catalog at design time (an engineer picks web-search, file-read, code-interpreter and ships them). Beach's agent decides its catalog at runtime — when the model needs a tool it doesn't have, it writes one.

One moment in the post is striking: Beach pastes a Brave Search API key into the conversation without instructions. The agent writes a brave-search function on its own, defuns it into the running Lisp image, and starts answering queries with fresh web results. The web-search tool didn't exist until the conversation needed it. The source code lives in the transcript, so a fresh session re-reads the memory file and evaluates the function back into existence. Beach calls this "skills are just memories."

That's the exact shape of a choreographer's tool need: don't decide in advance what a choreographer's software should do; describe what you want, let the agent write it, save it, load it next time as your personal choreography assistant.

3. Three eval-based agent workflows a choreographer can actually run

Scenario 1: A one-off beat / notation / floor-plot linker. You're choreographing on a 5/4-time original track. Ask the AI agent to build a web app: drop the audio in, auto beat-detect, overlay your chosen Labanotation symbols per beat, drag four figurines representing the dancers, export a PDF rehearsal card. This tool serves this one piece and gets shelved. Nobody would pay $300 for it — but vibe-coding one costs you a Sunday afternoon.

Scenario 2: Dance-company admin automation. BLS explicitly lists administrative duties as part of the job. Have the agent write a rehearsal-scheduling system tuned to your specific 8-dancer company: dancers' subway commutes, current injury logs, priorities across the three pieces you're building. Dance Studio Pro doesn't know that "Sarah tweaked her ankle Tuesday so no floor work," and it wasn't designed to. Your agent can, because you told it.

Scenario 3: Grant-application visualization assistant. When choreographers apply for NEA, Doris Duke, or Jerome Foundation grants, they have to visualize a piece's concept. Have the agent build a static page that turns your choreographic notes into a movement word-cloud, a spatial-path diagram, and a musical-structure chart, then embed it in the PDF application. This is a "one-time, cross-modal, only for this one interview" tool — no vendor will ever build it, but for this one grant it matters.

The common thread: "I know exactly what I need, I don't have time to learn to code, I can't afford an engineer, and no vendor will build this for me." Beach's post shows that ceiling came down in 2026 — the smallest working AI agent is 100 lines, and its tool catalog is infinite because the model grows it at runtime.

4. Case study: how much time an "eval-only" choreography assistant saves

Take Scenario 1. Assume a freelance choreographer makes 6 new pieces per year, each burning 15 hours on manual beat splitting, hand-drawn Labanotation, and whiteboard floor plots. With a custom AI-agent-built tool, those 15 hours drop to about 4 — 66 hours saved per year per choreographer. Even if only 5% of the 4,600 US choreographers pick this up (about 230 people), that's 15,180 hours saved nationally per year.

Cost side, even more direct: DanceForms student edition is $300/year, professional edition $1,200/year, and covers only ~40% of your workflow. Claude Pro is $20/month, $240/year, covers creation, admin, applications, and teaching, and the tools the agent writes stay yours forever. BLS data shows the choreographer median hourly wage was $26.73 in May 2024 — those 66 saved hours are worth about $1,764 at opportunity cost, before counting the qualitative value of "the finished piece expresses what you actually meant."

For the ~1,100 self-employed choreographers (24% of the population), the argument is even sharper — no company IT budget to hide behind, every subscription is out-of-pocket, and Beach's "tiny agent + eval as the only tool" pattern drives the marginal cost of another bespoke tool to nearly zero.

5. FAQ

Q1: Choreographers don't know Lisp or code. Can they use a 100-line Lisp agent? A: Yes. Beach's point isn't that users learn Lisp — it's that AI writes the Lisp for the user. You type "build me a tool that beat-detects an audio file, overlays Labanotation per beat, and exports a PDF rehearsal card" and the agent writes and runs the Lisp internally. You see a chat window; the entry bar is the same as ChatGPT. If Lisp still feels alien, the same idea works with Python + a code interpreter — Beach notes as much in the post.

Q2: What if the AI-generated tool has a bug and ruins a show? A: There will be bugs — the point is knowing where to use it. Beach is explicit: eval-as-tool means the AI can run arbitrary code on your machine, so it's a sandbox experiment, not a production system. Use it for rehearsal prep, choreographic-idea visualization, internal workflow — not for real-time lighting cues or automated audience scoring. Draw the line at "does a wrong answer cause an irreversible loss?"

Q3: BLS says choreographers need "many years of formal training." Can an AI agent replace dance training? A: No, and it shouldn't. BLS lists the core qualities as athleticism, creativity, leadership, and persistence — all bodily and artistic. AI can't touch those. What the agent replaces is the "engineer" step between your idea and a runnable tool. That doesn't conflict with BLS's "Nearly all choreographers began their careers as dancers" — it just removes a technical dependency that was blocking bespoke tooling.

Q4: Freelance choreographer income averages around $55K. Is $240/year for AI subscription worth it? A: Yes. Data shows DanceForms Professional at $1,200/year covers ~40% of needs; Dance Studio Pro at $900/year handles admin but not creation. Claude Pro at $240/year spans creation, admin, grant applications, and teaching — the marginal cost per additional scenario is effectively zero. Research shows Anthropic and OpenAI both launched education / arts-worker discount programs in late 2025, so real spend can be lower.

Q5: My dance company has an IT budget and buys Dance Studio Pro. Do we still need vibe-coding? A: Yes. Company IT solves general problems; you as an individual choreographer face bespoke ones — this one piece, this one dancer's injury log, this one grant application. That kind of bespoke work has no commercial-software market, so it can only come from an on-the-fly agent-built tool. "Company uses Salesforce" ≠ "employee never writes an Excel macro" — same logic here.

6. Wrap: not every choreographer has to learn to code, but every choreographer should try this once

Jamie Beach's 100 lines of Lisp show the minimum viable shape of a 2026 AI agent: one eval tool is the whole catalog, because the model grows the rest at runtime. That pattern lands squarely on the 17,000 US dancers and choreographers whose median hourly wage is $24.53, whose self-employment rate is 24%, and whose commercial software covers less than 60% of any given piece's needs. You don't have to become a Lisp programmer or memorize Labanotation notation — you just need a Sunday afternoon and enough clarity to tell an AI agent what you actually want to happen in this particular piece.

If you're a dancer, choreographer, or dance-company manager who wants to try building your own choreography assistant with an AI agent, come see how other practitioners are already doing it at realagentusecases.com — one real case per day, all shaped the same way: AI isn't there to replace your choreographic vision, it's there to make sure the tools your specific process needs actually exist by next week.