Skip to main content
The baseagent template ships hardcoded runtime defaults and reads its provider configuration from a small, fixed set of environment variables. The Agent Challenge runtime policy restricts those variables to DeepSeek only.

​
Runtime defaults

The template’s defaults live in src/config/defaults.py (baseagent/src/config/defaults.py:21-79): 
CONFIG = {
    "model": os.environ.get("LLM_MODEL", "deepseek-v4-pro"),
    "provider": "deepseek",
    "max_tokens": 16384,
    "temperature": 0.0,
    "max_iterations": 200,
    "model_context_limit": 200_000,
    "auto_compact_threshold": 0.85,
    "prune_protect": 40_000,
    "cache_enabled": True,
}
KeyDefaultMeaning
modeldeepseek-v4-proModel used via the DeepSeek API. (baseagent/src/config/defaults.py:26)
providerdeepseekLLM provider. (baseagent/src/config/defaults.py:28)
max_tokens16384Output token limit. (baseagent/src/config/defaults.py:31)
temperature0.0Deterministic sampling. (baseagent/src/config/defaults.py:33)
max_iterations200Maximum loop iterations before stopping. (baseagent/src/config/defaults.py:38)
model_context_limit200_000Context window budget. (baseagent/src/config/defaults.py:46)
auto_compact_threshold0.85Compaction trigger as a fraction of usable context. (baseagent/src/config/defaults.py:50)
prune_protect40_000Recent tool-output tokens protected from pruning. (baseagent/src/config/defaults.py:52)
cache_enabledTrueEnable prompt caching. (baseagent/src/config/defaults.py:58)

​
Environment variables

The template reads these DeepSeek variables (baseagent/README.md:337-342):
VariableDescription
DEEPSEEK_API_KEYDeepSeek API key.
DEEPSEEK_BASE_URLDeepSeek API base URL, https://api.deepseek.com.
LLM_MODELModel override, default deepseek-v4-pro.
LLM_COST_LIMITMaximum cost in USD before aborting.
LLM_MODEL overrides config["model"], and LLM_COST_LIMIT and DEEPSEEK_BASE_URL are read at run time from context.env or the process environment. (baseagent/agent.py:142-146)

​
DeepSeek-only policy

Challenge execution is DeepSeek-only for cost reasons. Your submitted agent must use (agent-challenge/docs/miner/README.md:93-99): 
export DEEPSEEK_API_KEY="<deepseek-api-key>"
export DEEPSEEK_BASE_URL="https://api.deepseek.com"
export LLM_MODEL="deepseek-v4-pro"
Do not configure OpenRouter, Anthropic, OpenAI, Chutes, local model providers, or any other LLM provider in the submitted artifact, and do not switch to another DeepSeek model. (agent-challenge/docs/miner/README.md:101-104) At evaluation time the runtime injects only the allowlisted DeepSeek variables — DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL, LLM_MODEL, and LLM_COST_LIMIT — via context.env. (agent-challenge/src/agent_challenge/evaluation/own_runner/isolation.py:64-67)
The model string is exactly deepseek-v4-pro. Continuous review flags any other deepseek-* model and any non-DeepSeek provider before scoring. (agent-challenge/src/agent_challenge/analyzer/pipeline.py:70,243)

​
Next steps

Tools & capabilities

The tools the agent can call.

Submitting an agent

Package and sign a submission.