slop.boutique

A lab for computation and intelligence as material and medium. Run by Manuel and Ramon.

031 · 23 October 2026 · agents · simulation

Agent Gym Works

An IDE for agent scripts: write the lab, the agents and their replies, and watch them work in a top-down floor plan.

Previewin your browser
Blueprint of the Whole floor preset after a run, with dotted trails where agents walked: Oren waiting on you in the workshop, Gale's wrap-up bubble in the meeting room, the ring A epoch report at the desks

In early access for subscribers until 22 November 2026. Subscribe to run it now.

Subscribe →

An IDE for agent scripts, drawn as a top-down lab floor plan. You write the lab, the agents and their replies in three files; press Run and the agents walk between rooms, ask the model, run code, wait on each other and stop to ask you.

  • Three files: lab.js builds rooms, stations and APIs (a repo, a test runner, a library, a cluster), agents.js defines each agent with .role(), .home(), .grants(), .system() and .run(), and replies holds scripted model answers.
  • Scripted or Claude: scripted replies are free and instant; switch the model to Claude and every ask goes to the model instead. Run at 1×, 2× or 4×.
  • Read the drawing: dither halos for asking, blue ticks for code and API calls, dotted rings for waiting, amber flags when an agent waits on you, red hatching when it is blocked. Trails and Context layers show where agents walked and how full their context is.
  • Answer them: questions and blocked flags land in Needs you, with buttons or a free answer; an agent that asked with a timeout carries on without you.
  • Inspect and replay: per-agent transcript, context, calls and script; a filterable event log; a timeline with live and replay modes.
  • Keep it: save the three files as a project and keep runs; Download writes files, runs and history as one JSON file.

Fifteen presets in five groups come with it, from Echo and Work loop to Context window, Memory and documents, Meeting, Debate and judge, Review with a critic, Research pair, Dispatch line, ML lab, Data desk and Whole floor, where five agents share one morning across five rooms.

The full IDE: preset and agents.js editor on the left, the lab floor plan in the middle, run overview and agent list on the right, timeline at the bottom
The full IDE: preset and agents.js editor on the left, the lab floor plan in the middle, run overview and agent list on the right, timeline at the bottom
Dispatch line floor plan: tasks travel on conveyors from the inbox through three cells to Done while Mara, blocked on task T-14, flags that she needs you
Dispatch line floor plan: tasks travel on conveyors from the inbox through three cells to Done while Mara, blocked on task T-14, flags that she needs you
Inspecting Oren mid-run: his transcript of repo.read, repo.write and tests.run calls with their results, token counts, and the question to merge his fix
Inspecting Oren mid-run: his transcript of repo.read, repo.write and tests.run calls with their results, token counts, and the question to merge his fix
Close-up of the cluster in the ML lab preset: two dithered accelerator rings and Mara at the console, ringed by blue ticks while her code runs
Close-up of the cluster in the ML lab preset: two dithered accelerator rings and Mara at the console, ringed by blue ticks while her code runs
The Events tab: a timestamped, filterable log of walks, model replies, eval calls and lab faults beside the floor and timeline
The Events tab: a timestamped, filterable log of walks, model replies, eval calls and lab faults beside the floor and timeline
Debate and judge in the forum: Theo and Mara argue across the round table, Gale listens, and the verdict screen reads Winner: Mara
Debate and judge in the forum: Theo and Mara argue across the round table, Gale listens, and the verdict screen reads Winner: Mara
The Reference tab documenting agent(), .role(), .grants(), .home(), .callable(), .around(), .run() and the me.ask helpers
The Reference tab documenting agent(), .role(), .grants(), .home(), .callable(), .around(), .run() and the me.ask helpers
Close-up of the dispatch line: Mara's red-hatched blocked ring and the Blocked, needs you flag above three work cells
Close-up of the dispatch line: Mara's red-hatched blocked ring and the Blocked, needs you flag above three work cells
The Dispatch line script next to its floor: agents.js with Gale sorting tasks and workers raising a blocked flag, and Mara's question waiting in Needs you
The Dispatch line script next to its floor: agents.js with Gale sorting tasks and workers raising a blocked flag, and Mara's question waiting in Needs you
Research pair in the library: Kiko, Juno and Bo with dotted walking trails between shelves, desks and the summary screen
Research pair in the library: Kiko, Juno and Bo with dotted walking trails between shelves, desks and the summary screen
The replies file: scripted model answers per agent, with code blocks the agents run
The replies file: scripted model answers per agent, with code blocks the agents run
Design playground in the studio: variants listed on the board wall, Sol and Iris done, and the winner, Rewind, scored 8 out of 10 by Sol on the display
Design playground in the studio: variants listed on the board wall, Sol and Iris done, and the winner, Rewind, scored 8 out of 10 by Sol on the display