Build & Finetune Agents

Framework Agnostic

Weaver is deeply optimized for the NexAU and NexRL workflow, providing seamless integration and best practices for building and finetuning agentic models. That said, Weaver remains framework-agnostic and can be integrated with other agent frameworks to suit your needs.

NexAU Quick Start: Build a Web Research Agent

NexAU is a general-purpose agent framework for building tool-using LLM agents. It provides declarative configuration, flexible tool systems, and seamless tracing for both standalone usage and reinforcement learning.

We're going to build a web research agent based on nexau/examples/deep_research/:

type: agent
name: deep_research_agent
max_context_tokens: 100000

system_prompt: |
  Date: {{date}}
  You are a research agent. Search for information using web_search and web_read.

system_prompt_type: jinja
tool_call_mode: openai

llm_config:
  temperature: 0.7
  max_tokens: 4096
  api_type: openai_chat_completion

tools:
  - name: web_search
    yaml_path: ./tools/WebSearch.yaml
    binding: nexau.archs.tool.builtin.web_tool:web_search
  - name: web_read
    yaml_path: ./tools/WebRead.yaml
    binding: nexau.archs.tool.builtin.web_tool:web_read

Run it:

from nexau import Agent, AgentConfig, LLMConfig
from datetime import datetime
import os

agent_config = AgentConfig.from_yaml("deep_research_agent.yaml")
agent_config.llm_config = LLMConfig(
    model=os.getenv("LLM_MODEL"),
    base_url=os.getenv("LLM_BASE_URL"),
    api_key=os.getenv("LLM_API_KEY")
)

agent = Agent(agent_config)
response = agent.run(
    "What are the latest developments in quantum computing?",
    context={"date": datetime.now().strftime("%Y-%m-%d")}
)
print(response)

That's all you need to build a web research agent. For more advanced features, e.g. sub-agents, skills and middlewares, etc., please refer to NexAU's doc.

Finetune Agents using NexRL

NexAU agents can be trained with reinforcement learning via NexRL.

Agent Workspace Setup

For NexRL integration, organize your agent in a workspace:

recipe/my_task/agent_workspace/
├── agent_config.yaml       # NexAU agent definition
├── evaluator.py           # Reward computation
├── custom_worker.py       # Optional: custom formatting
└── tools/                 # Tool implementations

Agent Configuration for RL

Enable tracing for NexRL:

llm_config:
  temperature: 0.7
  max_tokens: 8192
  api_type: openai_chat_completion

  # RL-specific settings (optional)
  logprobs: true
  extra_body:
    train_mode: true
    include_stop_str_in_output: true
    skip_special_tokens: false
    return_tokens_as_token_ids: true

# Required for NexRL
tracers:
  - import: nexau.archs.tracer.adapters.in_memory:InMemoryTracer

NexRL Integration

In your NexRL recipe, reference the agent:

rollout_worker:
  type: "nexau"
  nexau_agent_config_path: "recipe/my_task/agent_workspace/agent_config.yaml"
  evaluator_module_path: "recipe/my_task/agent_workspace/evaluator.py:MyEvaluator"
  nexau_agent_workspace: "recipe/my_task/agent_workspace"
  task_name: "my_task"

Example Recipe: NexRL/recipe/weaver_nexau_deepsearch_qwen3_8b/

For complete details on NexRL integration, see the NexRL guide.

Next Steps

• Full documentation: NexAU GitHub
• Advanced features: Skills, MCP integration, context compaction
• RL training: NexRL guide for reinforcement learning with Weaver
• Examples: Check NexAU Examples for complete agent implementations