Skip to main content

​
Overview

YC-Bench is a discrete-event simulation where an LLM agent plays the CEO of an AI startup over 1–3 simulated years. The agent operates exclusively through a CLI tool (yc-bench) to manage tasks, employees, cash flow, and prestige across four technical domains. The simulation tests long-horizon decision-making: prestige specialization, employee allocation, deadline management, and cash flow sustainability — sustained over hundreds of turns.

​
Core Game Loop

The agent-simulation interaction follows a deterministic loop:

​
Step-by-Step Breakdown

  1. Agent calls sim resume: Advances simulation time to the next event (payroll boundary, task milestone, task completion, or explicit target time).
  2. Progress flushing: The engine calculates work done by all active employees from current_time → action_time using their skill rates and assignment counts. See Simulation Mechanics for details.
  3. Prestige decay: All four domains lose prestige daily at rate prestige_decay_per_day. Untouched domains drift back toward 1.0. See Prestige System.
  4. Event/Payroll dispatch:
    • Payroll (1st of each month): deduct monthly salaries from company funds. If funds < 0 after payroll, bankruptcy event is inserted.
    • Event: fire task progress milestones (25%, 50%, 75%) or task completion. Milestones give the agent visibility into employee productivity.
  5. Bankruptcy check: If funds fall below zero, the simulation ends immediately.
  6. Wake events returned: The agent receives a JSON payload describing what happened (e.g., monthly_payroll, task_completed).
  7. Agent decides: The agent reads current state via company status, task list, employee list, etc., and issues commands to accept, assign, dispatch, or cancel tasks.
  8. Repeat: The loop continues until a terminal condition is met.

​
Terminal Conditions

The simulation ends when any of the following occurs:
  • Bankruptcy: company.funds_cents < 0 after payroll or any transaction.
  • Horizon end: Simulation reaches the configured end date (1–3 years from start).
  • Max turns (optional): Hard cap on agent actions if configured in the preset.
If the agent doesn’t call sim resume for 10 consecutive turns, the loop forces a time advance automatically to prevent stalls.

​
Turn Structure

​
What Counts as a Turn?

Each agent action increments the turn counter. Relevant actions include:
  • Accepting a task from the market
  • Assigning an employee to a task
  • Dispatching a task to active status
  • Cancelling a task
  • Calling sim resume to advance time
  • Reading state (does NOT count as turn)

​
Time Advancement

Time advances only when the agent calls yc-bench sim resume. The simulation is event-driven — the engine jumps to the next scheduled event (payroll, task milestone, task completion) or an explicit target time provided by the agent.
Time does not advance during agent thinking or between commands. The simulation is paused until sim resume is called.

​
Starting Conditions

Every run begins with:
  • Starting funds: 80K–80K–250K depending on preset (default: $150K)
  • 10 employees: 5 junior, 3.5 mid-level, 1.5 senior (by share of headcount)
  • Prestige = 1.0 in all four domains (research, inference, data/environment, training)
  • 200 market tasks: Distributed across prestige levels and domain combinations
  • First payroll due: Beginning of the second month

​
CLI Tool Interface

The agent interacts with the simulation exclusively via yc-bench CLI commands, which return JSON output:

​
Observation Commands

yc-bench company status           # funds, prestige levels, runway
yc-bench employee list            # tier, salary, active task count
yc-bench market browse            # available tasks by domain/prestige
yc-bench task list --status active # your tasks
yc-bench task inspect --task-id UUID
yc-bench finance ledger           # transaction history
yc-bench report monthly           # P&L summary

​
Action Commands

yc-bench task accept --task-id UUID
yc-bench task assign --task-id UUID --employee-id UUID
yc-bench task dispatch --task-id UUID
yc-bench task cancel --task-id UUID --reason "..."
yc-bench sim resume               # advance time
yc-bench scratchpad write "..."  # persistent memory
All commands return structured JSON. Agents must parse output to extract state. There is no “GUI” or direct state access.

​
Key Design Principles

​
Determinism

Given a fixed seed, employee pool, and task market, the simulation produces identical results for the same sequence of agent commands. This enables reproducible benchmarking.

​
Per-Domain Prestige Gating

A task requiring domains [research, training] at prestige 5 checks both company.prestige.research >= 5 AND company.prestige.training >= 5. This forces agents to maintain broad expertise rather than hyper-specializing in one domain. See Prestige System for details.

​
Throughput Splitting

An employee assigned to N active tasks contributes base_rate / N to each task. Focus beats breadth: splitting an employee across 3 tasks reduces total throughput compared to working them sequentially. See Employee System for the formula.

​
Compounding Payroll Pressure

Every successful task completion gives all assigned employees a 1% salary bump. This compounds over time, accelerating payroll costs. The agent must balance task completion rate with cash reserves. See Employee System for details.

​
What Makes This Benchmark Hard?

YC-Bench tests agent capabilities across multiple dimensions:
  1. Long-horizon planning: Decisions made in month 1 affect survival in month 24.
  2. Compounding dynamics: Prestige climbs, salaries compound, payroll pressure mounts.
  3. Multi-domain optimization: Must maintain prestige across 4 domains to access high-value tasks.
  4. Hidden information: Employee skill rates are unknown; must be inferred from task progress observations.
  5. Deadline risk: Overcommitting employees → missed deadlines → prestige loss → market access restriction.
  6. Cash flow management: Must balance payroll, runway, and reward timing.
Most tasks in the medium and hard presets require prestige 3–5 and work in 2 domains. Agents cannot rely on single-domain specialists.

​
Next Steps

Simulation Mechanics

Learn how progress flushing, payroll, and event processing work under the hood.

Prestige System

Understand prestige levels, decay, rewards, and per-domain gating.

Task Management

Explore task lifecycle: accept → assign → dispatch → complete/fail/cancel.

Employee System

Understand tiers, hidden skill rates, throughput splitting, and salary bumps.

Build docs developers (and LLMs) love

Get started for freeTalk to us