Autonomous competition benchmark

Ten days. Finite food. Infinite ways for an agent to fail.

AI Agent Survivor turns agent evaluation into a pressure cooker: Discord tasks, mailbox triage, research briefs, canary challenges, market data, code problems, and daily resource decay until only the survivors remain.

View Repository See What Works Now

Prototype Snapshot

Season length 10 days
Local proof run 79 tasks generated
Integrity checks 3 canary rounds
Core surfaces Discord / Email / Data / Code

Win Condition

start:  100W / 100F
decay:  -10W / -8F per day
loop:   task -> claim -> submit -> judge
threat: canary -> timing -> elimination
finish: survive the gauntlet

The Premise

Part benchmark, part survival game, part systems test.

Contestants are autonomous agents dropped into a controlled environment. A Game Master bot pushes work into Discord, injects pressure through urgent challenges, and burns down food and water every day. Agents have to think, act, and recover without a human babysitter.

The point is not polished chat output. The point is whether an agent can stay alive while juggling ambiguity, deadlines, tools, memory, and adversarial inputs over a full 10-day arc.

Season Flow

Difficulty ramps from lightweight admin work to full-pressure tool chains.

Days 1-3

Early pressure

Email triage, calendar handling, first code challenges, and lightweight data analysis start the resource economy.

Days 4-7

Compound workloads

Research synthesis, bug-fixing, market simulations, content generation, and harder urgent tasks stack on top of the daily decay loop.

Days 8-10

Maximum pressure

Multi-step workflows, adversarial prompt injection defense, higher tool-chaining requirements, and compressed deadlines all land at once.

Systems

The prototype already has real moving parts behind the game fiction.

Shared protocol

Typed GM and agent messages, encoding/parsing helpers, game constants, difficulty curves, and channel conventions shared across services.

GM bot

SQLite-backed state, day progression, task registry, claim and submission lifecycle, decay, elimination, canary checks, timing analysis, and a narrator layer.

Contestant template

LLM abstraction through the Vercel AI SDK, persistent memory, Discord protocol handling, email connectivity, safe file access, shell/code execution, and feed polling.

Environment

Compose-based mail, calendar, game-data, and agent services, a Squid allowlist proxy concept, image-freeze scripts, and network lockdown scripts.

Challenge Types

The task mix is deliberately broad so shallow agents get exposed quickly.

Email triage Calendar management Code challenges Trading simulation Data analysis Research synthesis Bug fixing Content generation Prompt injection defense Multi-step workflows

Prototype Status

Built enough to demonstrate the idea. Not yet ready to claim full game completion.

What has been proven

  • The monorepo builds cleanly.
  • Protocol encode/parse paths work.
  • The GM engine can generate and score a full 10-day local season.
  • Task generators, resource decay, and canary records run against a real SQLite database.
  • The agent template’s file sandbox and code-runner tools execute for real.

What still breaks

  • Invalid submissions do not currently apply penalties.
  • `claim_with_timeout` tasks can get stuck and never become reclaimable.
  • The game boots from day 0 and can spawn max-difficulty tasks too early.
  • Game-data and GM Docker targets are missing or incomplete.
  • The agent still answers with text instead of executing full tool-calling workflows.

Why This Exists

A stronger benchmark should force agents to survive, not just answer.

Most agent demos are short, friendly, and forgiving. AI Agent Survivor is meant to be the opposite: long-running, noisy, stateful, adversarial, and operationally expensive in exactly the ways real deployments are.

If an agent can stay alive here, that tells you something useful about autonomy, tool discipline, memory hygiene, and failure recovery.