Agent Kitchen

Replace prompt engineering with evolution.

Describe the agent you want. A population of competing AI agents is dropped into OASIS social simulations: negotiating deals, posting on feeds, responding to crises. An LLM judge scores them. The weak die. The survivors mutate. The best prompt emerges, battle-tested and never written by a human.

"best negotiator" → scenarios generated → agents spawned → simulations run → scored → weakest die → survivors mutate → repeat → evolved prompt out

The system auto-generates diverse scenarios (salary negotiation, vendor contract, public outreach on social media). A population of agents with different strategies competes across all of them inside live OASIS simulations. An LLM judge evaluates every interaction against a rubric. Bottom performers are eliminated. Survivors are mutated and crossed over to produce the next generation. After N cycles, the fittest agent's prompt is the output.

Most agent benchmarks test if an AI can generate good text. That's not enough. OASIS is a social simulation platform from CAMEL-AI that models realistic multi-agent environments at scale. Agent Kitchen uses it as the fitness arena, the world where agents prove themselves.

A simple LLM wrapper tests what an agent says. OASIS tests what an agent does: whether it builds influence, handles backlash, negotiates under pressure, and adapts when others push back.

Each agent is a structured prompt with six independently evolvable sections:

Mutation operators mirror real genetics: point mutations (50%), rewrites (20%), insertions (10%), deletions (10%), and crossover (10%). Tactics and style evolve fast. Role and constraints evolve slowly, just like biology. This means evolution explores the strategy space efficiently instead of randomly rewriting entire prompts.

Evolution is a black box if you can't watch it happen. The terminal dashboard (built with Bubble Tea + Lip Gloss) streams everything in real time:

• Input: describe your goal, optionally define a custom rubric, set generations and scenarios
• Grid View: each cell is a scenario running in parallel with live agent conversations
• Detail View: expand any cell, tab between agents to compare how each one handles the same situation
• Results: fitness charts, ranked leaderboard with per-scenario breakdowns, the full evolved prompt, and one-key export

The sidebar explains what's happening at every phase: generating scenarios, simulating interactions, scoring agents, natural selection, breeding the next generation.

Frontend: Go + Bubble Tea TUI. Reads JSONL events from the backend via a pipe and renders them in real time.

Backend: Python + OASIS. An orchestrator coordinates the full evolution loop: scenario generation, population seeding, parallel OASIS simulations, LLM evaluation, selection, and breeding. Custom SocialAgent subclasses override OASIS's default prompts for focused interactions in both private chat and public feed scenarios. All data persists in SQLite (one DB per scenario per generation).

git clone https://github.com/karmette/agent-kitchen.git
cd agent-kitchen

# Backend
cd backend && uv sync --no-dev
cp .env.example .env  # add your API key

# Frontend
cd ../frontend && go build -o agent-kitchen
./agent-kitchen

Requires Python 3.11+, Go 1.21+, and an OpenAI-compatible API key.

AI slop implies the existence of AI peak