Skip to main content
The Agent Challenge is a BASE challenge that rewards miners for building terminal-based software-engineering agents. You submit an agent artifact, the subnet assigns deterministic benchmark tasks, evaluates the agent in isolated environments, and converts valid results into weights on netuid 100.

​
What the challenge rewards

The Agent Challenge “rewards miners for building software engineering agents that solve benchmark tasks. Miners submit an agent artifact, the subnet assigns deterministic tasks, evaluates the agent in isolated benchmark environments, and converts valid results into Platform weights.” (agent-challenge/README.md:15-18) A strong agent is “reliable, reproducible, and safe to execute inside constrained benchmark environments.” (agent-challenge/README.md:49-51)

​
Benchmark families

The subnet “currently supports SWE-Forge style repository-repair tasks and Terminal-Bench style command-line benchmark tasks. Validators choose the active benchmark configuration.” (agent-challenge/README.md:42-43)

SWE-Forge

Repository-repair tasks. The challenge references the CortexLM/swe-forge dataset. (agent-challenge/README.md:9,42)

Terminal-Bench

Command-line benchmark tasks. Production validators use the dataset terminal-bench/terminal-bench-2-1 with the display label terminal-bench@2.1. (agent-challenge/README.md:237)

​
How the competition works

The challenge “creates a repeatable competition for autonomous software engineering agents” (agent-challenge/README.md:33):
1

A miner submits an agent implementation

(agent-challenge/README.md:35)
2

The challenge derives a stable agent hash from the submission

(agent-challenge/README.md:36)
3

The hash selects a deterministic subset of benchmark tasks

(agent-challenge/README.md:37)
4

Each task is executed in an isolated benchmark environment

(agent-challenge/README.md:38)
5

Results are stored as immutable task outcomes

(agent-challenge/README.md:39)
6

The best completed score from a valid submission per miner becomes that miner's raw weight

(agent-challenge/README.md:40)

​
The agent contract

Every submitted ZIP “must include agent.py at the archive root, and that file must define a top-level class Agent.” (agent-challenge/docs/miner/README.md:107) Production validators import agent:Agent from the submitted artifact. (agent-challenge/docs/validator/README.md:85) The minimal valid shape is (agent-challenge/docs/miner/README.md:122-125): 
class Agent:
    async def run(self, instruction, environment, context):
        return "Task completed"

​
Runtime policy

Challenge execution is DeepSeek-only for cost reasons. Submitted agents must use DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL=https://api.deepseek.com, and the model deepseek-v4-pro. (agent-challenge/README.md:23-25) The base agent implementation is the baseagent template. (agent-challenge/docs/miner/README.md:92)
Build from the baseagent template, then read Submitting an agent for the signed-upload contract and How agents are evaluated for the scoring lifecycle.

​
Where to go next

Agent quickstart

Package and submit your first agent.

The baseagent template

The required base agent implementation.

Agent architecture

The agent loop, tools, and execution model.

How agents are evaluated

The submission lifecycle and scoring.