TokenGolf

Created by team OMEGA on July 07, 2026
AgentOps
Hybrid Token-Efficient Routing Agent

router-agent answers all eight Track-1 categories while minimizing the Fireworks tokens the leaderboard scores. Its key optimization: both allowed Fireworks models (minimax-m3, kimi-k2p7-code) are reasoning models whose hidden reasoning dominates token cost, so the agent disables it (reasoning_effort="none"), measured to cut tokens ~46% with no accuracy loss. It implements the harness contract exactly (/input/tasks.json → /output/results.json, all inference through FIREWORKS_BASE_URL, models read from ALLOWED_MODELS), builds linux/amd64, and stays well under the 10 GB / 60 s / 30 s-per-request limits.

Category tags:
GithubPresentationDemo
Explore more applications

Kalari: Contained Rehearsals for API Upgrades

Kalari plugs into your coding agent and rehearses API upgrades in throwaway microVMs. It proves what changed, contains every run, and records what the run proved. Measured: 4.2x cheaper and 3.3x faster than the baseline.

Kalari

VULTRDeepSeek V3

First Answer Agent: AI Front Desk

An AI front desk for small businesses. No waiting on hold, voicemail or callbacks: customers get an answer on the page right away, any hour. It offers open times, books the visit in the app and brings in a person when needed. Runs on Vultr.

First answer

VULTR

Blast Radius Zero

Web agent that generates Python via Vultr Serverless Inference and runs it in a disposable Docker sandbox (--network none, memory/CPU caps, --rm) on a Vultr VM.

kbirdg-sandbox

VULTR

JARVIS ProofOps

A working commerce QA prototype on Vultr that authorizes browser actions, executes under gVisor, and independently checks outcomes with retained evidence.

JARVIS ProofOps

VULTR

SandboxOps: Blast Radius Zero on Vultr Compute

Blast-radius-zero agent ops on Vultr: Docker sandbox on Cloud Compute, verify, SHA-256 evidence. Inference (glm-5.3) plans; Compute contains. Live http://64.177.13.126/ · suite 3/3+5/5

SandboxOps

AgentOpsVULTR