Hermes Agent sub-agent delegation and batch processing

This chapter introduces Hermes' two parallel processing capabilities: sub-agent delegation, which allows a main agent to distribute tasks to multiple sub-agents for parallel execution, and a batch processing system that can run agents in parallel on hundreds of prompts and generate structured training data.

Sub-agent delegation and parallelism

Hermes supports delegating tasks from the main Agent to sub-Agents, enabling parallel workflows.

Sub-agents viadelegationToolset implementation:

Example

# File path: ~/.hermes/config.yaml
# Ensure the delegation toolset is enabled
toolsets:
- delegation

Main agent usesspawn_agentorspawn_parallelTool-derived Agent.

The following pseudocode illustrates the mechanism of parallel delegation:

Example

# The parallel delegation mechanism of Agent in sessions (pseudocode, explaining the principle)
# spawn_parallel dispatches multiple sub-agents simultaneously, each executing independently.

spawn_parallel([
{"prompt": "Analyze Q1 sales data and generate trend report"},
{"prompt": "Analyze Q2 sales data and generate trend report"},
{"prompt": "Analyze Q3 sales data and generate trend report"}
])
# Three sub-Agents work in parallel, with results aggregated and returned to the main Agent

Background task

Start background tasks directly in the session:

Example

# Start Background Task in Session
/background check all servers and report downtime

Kanban Multi-Agent Collaboration

The Kanban toolset provides more structured multi-Agent collaboration — an orchestrator Agent distributes tasks to the board, and multiple worker Agents claim and execute them:

Example

# Enable kanban toolset (requires explicit enabling; all/* will not automatically enable it)
hermes tools enable kanban

# Add role descriptions to Profile so the orchestrator knows each Agent's capabilities.
hermes profile create researcher \
--description "Skilled at reading source code and external documentation, producing research reports"
hermes profile create coder \
--description "Skilled at implementing features, fixing bugs, and writing tests"

The Kanban toolset must be explicitly enabled — even all or * will not automatically turn it on. This is intentional design, because multi-Agent collaboration consumes more tokens.

Batch Processing and RL Training Data Generation

Batch processing lets you run agents in parallel on hundreds or thousands of prompts, generating structured trajectory data—mainly used for model fine-tuning training data generation and evaluation.

Quick start

Example

# Prepare a Prompt dataset in JSONL format
cat data/prompts.jsonl
# {"prompt": "Write a Python function to find the longest palindromic substring"}
# {"prompt": "Create user authentication REST API endpoints with Flask"}
# {"prompt": "DebugThisError:TypeError: cannot unpack non-iterable"}

# Run Batch Processing
python batch_runner.py \
--dataset_file=data/prompts.jsonl \
--batch_size=10 \
--run_name=my_first_run \
--model=anthropic/claude-sonnet-4-6 \
--num_workers=4

# Resume interrupted task from checkpoint
python batch_runner.py \
--dataset_file=data/prompts.jsonl \
--batch_size=10 \
--run_name=my_first_run \
--resume

Dataset format

Each line is a JSON object, must containpromptField, optionally contains additional context:

Example

{"prompt": "Write a Python function to find palindromes"}
{"prompt": "DebugThisError", "cwd": "/home/user/project", "docker_image": "python:3.12"}

Main parameters

Parameter	Default value	Description
--dataset_file	Required	JSONL dataset path
--batch_size	Required	Number of prompts per batch
--run_name	Required	Run name (used for output directory and checkpoint resumption)
--model	claude-sonnet-4.6	Models used
--num_workers	4	Number of parallel worker processes
--max_turns	10	Maximum tool call rounds per prompt
--distribution	"default"	Toolset distribution (randomly sampling different tool combinations)
--reasoning_effort	—	Reasoning effort: none/minimal/low/medium/high/xhigh
--resume	false	Resume from checkpoint

Output structure

Batch processing generates the following directory structure:

data/my_first_run/
├── trajectories.jsonl    # 最终合并输出（所有批次）
├── batch_0.jsonl         # 各批次结果
├── batch_1.jsonl
├── checkpoint.json       # 断点续传记录
└── statistics.json       # 工具使用统计

Each trajectory is output in ShareGPT format, including full conversation history, tool invocation records, and reasoning coverage statistics:

Example

{
"prompt_index": 42,
"conversations": [
{"from": "human", "value": "Write a function to find palindromes..."},
{"from": "gpt", "value": "I'll create this function...",}
"tool_calls": [{"name": "terminal", "arguments": {...}}]},
{"from": "tool", "value": "...execution result..."},
{"from": "gpt", "value": "This is the completed function..."}
],
"completed": true,
"toolsets_used": ["terminal", "file"],
"tool_stats": {
"terminal": {"count": 2, "success": 2, "failure": 0}
}
}

Quality filtering

After the batch run ends, two filters are automatically applied:

filter item	Rules	Purpose
Zero inference filtering	Samples without any trace of reasoning were discarded.	Ensure the output contains the thinking process to improve training data quality
Hallucinated tool name filtering	Entries containing tool calls that are not in the valid tool list are filtered out	Exclude invalid tool calls produced by model hallucination

Reasoning traces refer to <REASONING_SCRATCHPAD> markers or native thinking tokens. The filtered data is more suitable for model fine-tuning.

Integration with the Atropos RL environment

Batch processing can directly generate Atropos-compatible training data:

Example

# Generate Atropos-compatible training data
python batch_runner.py \
--dataset_file=data/coding_tasks.jsonl \
--run_name=atropos_run \
--model=nous-hermes-3.1 \
--num_workers=8 \
--reasoning_effort=high

# The output trajectories.jsonl can be directly used as Atropos training input

other extensions