`HierarchicalSwarm`¶

The HierarchicalSwarm is a sophisticated multi-agent orchestration system that implements a hierarchical workflow pattern. It consists of a director agent that coordinates and distributes tasks to specialized worker agents, creating a structured approach to complex problem-solving.

graph TD
    A[Task] --> B[Director]
    B --> C[Plan & Orders]
    C --> D[Agents]
    D --> E[Results]
    E --> F{agent_as_judge?}
    F -->|Yes| G[Judge Agent — scores outputs]
    F -->|No| H{director_feedback_on?}
    H -->|Yes| I[Director Feedback]
    H -->|No| J{More Loops?}
    G --> J
    I --> J
    J -->|Yes| B
    J -->|No| K[Output]

The Hierarchical Swarm follows a clear workflow pattern:

Task Reception: User provides a task to the swarm
Planning: Director creates a comprehensive plan and distributes orders to agents
Execution: Individual agents execute their assigned tasks sequentially or in parallel
Evaluation: Optionally, a judge agent scores each output with structured numeric scores
Feedback Loop: Director evaluates results and issues new orders if needed (up to max_loops)
Context Preservation: All conversation history and context is maintained throughout the process

Key Features¶

Feature	Description
Hierarchical Coordination	Director agent orchestrates all operations
Specialized Agents	Each agent has specific expertise and responsibilities
Iterative Refinement	Multiple feedback loops for improved results
Context Preservation	Full conversation history maintained
Flexible Output Formats	Support for various output types (dict, str, list)
Comprehensive Logging	Detailed logging for debugging and monitoring
Live Streaming	Real-time streaming callbacks for monitoring agent outputs
Token-by-Token Updates	Watch text formation in real-time as agents generate responses
Hierarchy Visualization	Visual tree representation of swarm structure with `display_hierarchy()`
Interactive Dashboard	Real-time Hierarchical Swarms dashboard for monitoring swarm operations
Advanced Planning	Optional planning phase before task distribution for better coordination
Parallel Execution	Run all agents concurrently via `ThreadPoolExecutor` with `parallel_execution=True`
Agent-as-Judge	Structured per-agent scoring (0–10) after each cycle with `agent_as_judge=True`
Async Entry Point	Non-blocking `arun()` method wraps `run()` in `asyncio.to_thread()`

Constructor¶

`HierarchicalSwarm.init()`¶

Initializes a new HierarchicalSwarm instance.

Parameters¶

Parameter	Type	Default	Required	Description
`agents`	`AgentListType`	`None`	Yes	List of worker agents in the swarm. Must not be empty
`name`	`str`	`"HierarchicalAgentSwarm"`	No	The name identifier for this swarm instance
`description`	`str`	`"Distributed task swarm"`	No	A description of the swarm's purpose and capabilities
`director`	`Optional[Union[Agent, Callable, Any]]`	`None`	No	The director agent that orchestrates tasks. If None, a default director will be created
`max_loops`	`int`	`1`	No	Maximum number of feedback loops between director and agents (must be > 0)
`output_type`	`OutputType`	`"dict-all-except-first"`	No	Format for output (dict, str, list)
`director_model_name`	`str`	`"gpt-5.4"`	No	Model name for the main director agent
`director_name`	`str`	`"Director"`	No	Name identifier for the director agent
`director_temperature`	`float`	`0.7`	No	Temperature setting for the director agent (controls randomness)
`director_top_p`	`float`	`0.9`	No	Top-p (nucleus) sampling parameter for the director agent
`director_system_prompt`	`str`	`HIEARCHICAL_SWARM_SYSTEM_PROMPT`	No	System prompt for the director agent
`director_feedback_on`	`bool`	`True`	No	Whether director freeform feedback is enabled after each cycle
`feedback_director_model_name`	`str`	`"gpt-5.4"`	No	Model name for the feedback director agent
`add_collaboration_prompt`	`bool`	`True`	No	Whether to add collaboration prompts to worker agents
`multi_agent_prompt_improvements`	`bool`	`False`	No	Enable enhanced multi-agent collaboration prompts
`interactive`	`bool`	`False`	No	Enable interactive mode with Hierarchical Swarms dashboard visualization
`planning_enabled`	`bool`	`True`	No	Enable planning phase before task distribution
`autosave`	`bool`	`True`	No	Whether to enable autosaving of conversation history
`verbose`	`bool`	`False`	No	Whether to enable verbose logging
`parallel_execution`	`bool`	`False`	No	Run all agents concurrently using `ThreadPoolExecutor`. Outputs are written to conversation in original submission order
`agent_as_judge`	`bool`	`False`	No	After each cycle, spin up a judge agent that scores every worker output (0–10) using a structured `JudgeReport`. Takes priority over `director_feedback_on`
`judge_agent_model_name`	`str`	`"gpt-5.4"`	No	Model used by the judge agent when `agent_as_judge=True`

Returns¶

Type	Description
`HierarchicalSwarm`	A new HierarchicalSwarm instance

Raises¶

Exception	Condition
`ValueError`	If no agents are provided or `max_loops` is invalid

Core Methods¶

`run()`¶

Executes the hierarchical swarm for a specified number of feedback loops, processing the task through multiple iterations for refinement and improvement.

Parameters¶

Parameter	Type	Default	Required	Description
`task`	`Optional[str]`	`None`	Yes*	The initial task to be processed by the swarm. If None and interactive mode is enabled, will prompt for input
`img`	`Optional[str]`	`None`	No	Optional image input for the agents
`streaming_callback`	`Optional[Callable[[str, str, bool], None]]`	`None`	No	Callback for real-time streaming. Parameters: `(agent_name, chunk, is_final)`

*Required if interactive=False

Returns¶

Type	Description
`Any`	The formatted conversation history, formatted according to `output_type`

Example¶

from swarms import Agent, HierarchicalSwarm

research_agent = Agent(
    agent_name="Research-Specialist",
    agent_description="Expert in market research and analysis",
    model_name="gpt-5.4",
    max_loops=1,
)

financial_agent = Agent(
    agent_name="Financial-Analyst",
    agent_description="Specialist in financial analysis and valuation",
    model_name="gpt-5.4",
    max_loops=1,
)

swarm = HierarchicalSwarm(
    name="Financial-Analysis-Swarm",
    agents=[research_agent, financial_agent],
    max_loops=2,
    director_model_name="gpt-5.4",
)

result = swarm.run(task="Analyze the market potential for Tesla (TSLA) stock")
print(result)

`arun()`¶

Async entry point that wraps run() in asyncio.to_thread(), allowing the swarm to be awaited without blocking the event loop. Accepts the same parameters as run().

Parameters¶

Parameter	Type	Default	Required	Description
`task`	`Optional[str]`	`None`	Yes*	The task to be processed by the swarm
`img`	`Optional[str]`	`None`	No	Optional image input
`streaming_callback`	`Optional[Callable[[str, str, bool], None]]`	`None`	No	Callback for real-time streaming

Returns¶

Type	Description
`Any`	Same result as `run()`

Example¶

import asyncio
from swarms import Agent, HierarchicalSwarm

agents = [
    Agent(agent_name="Researcher", model_name="gpt-5.4", max_loops=1),
    Agent(agent_name="Analyst", model_name="gpt-5.4", max_loops=1),
]

swarm = HierarchicalSwarm(
    name="Async-Swarm",
    agents=agents,
    director_model_name="gpt-5.4",
)

result = asyncio.run(swarm.arun(task="Summarize recent AI research trends"))
print(result)

Use arun() when integrating the swarm into an existing async application (e.g., FastAPI, async CLI tools) so the swarm does not block the event loop.

`batched_run()`¶

Executes the swarm for multiple tasks in sequence, running the complete workflow independently for each task.

Parameters¶

Parameter	Type	Default	Required	Description
`tasks`	`List[str]`	-	Yes	List of tasks to be processed
`img`	`Optional[str]`	`None`	No	Optional image input for the tasks
`streaming_callback`	`Optional[Callable[[str, str, bool], None]]`	`None`	No	Callback for streaming agent outputs

Returns¶

Type	Description
`List[Any]`	List of results for each processed task

Example¶

from swarms import Agent, HierarchicalSwarm

swarm = HierarchicalSwarm(
    name="Analysis-Swarm",
    agents=[
        Agent(agent_name="Market-Analyst", model_name="gpt-4.1", max_loops=1),
        Agent(agent_name="Technical-Analyst", model_name="gpt-4.1", max_loops=1),
    ],
    director_model_name="gpt-5.4",
)

results = swarm.batched_run(tasks=[
    "Analyze Apple (AAPL) stock performance",
    "Evaluate Microsoft (MSFT) market position",
    "Assess Google (GOOGL) competitive landscape",
])
for i, result in enumerate(results):
    print(f"Task {i+1}:", result)

`display_hierarchy()`¶

Prints a visual tree representation of the swarm structure — director at the top, worker agents as branches — using Rich formatting.

Example¶

swarm.display_hierarchy()

┌─ HierarchicalSwarm Hierarchy: My Swarm ─┐
│ 🎯 Director [gpt-4o-mini]               │
│ ├─ 🤖 Research-Analyst [gpt-4o-mini]   │
│ ├─ 🤖 Data-Analyst [gpt-4o-mini]       │
│ └─ 🤖 Strategy-Consultant [gpt-4o-mini]│
└─────────────────────────────────────────┘

New Optional Features¶

Parallel Execution (`parallel_execution`)¶

By default, agents run sequentially — each agent waits for the previous one to finish before starting. With parallel_execution=True, all agents in a cycle are submitted to a ThreadPoolExecutor simultaneously and run concurrently.

How it works:

All call_single_agent() calls are submitted to a thread pool in one loop.
as_completed() collects results as each agent finishes (fastest first).
After all futures resolve, outputs are written to conversation in the original submission order to guarantee deterministic history.

Worker count is set to max(1, int(os.cpu_count() * 0.75)) — 75% of available CPU cores.

When to use it: - Agents work independently with no dependency on each other's output - You want to minimize wall-clock time on large agent pools

When to avoid it: - Agents need to read each other's outputs (use sequential mode so conversation context is populated in order) - Tasks require strict ordering of agent contributions

from swarms import Agent, HierarchicalSwarm

swarm = HierarchicalSwarm(
    name="Parallel-Swarm",
    agents=[
        Agent(agent_name="Research-Analyst", model_name="gpt-5.4", max_loops=1),
        Agent(agent_name="Data-Analyst", model_name="gpt-5.4", max_loops=1),
        Agent(agent_name="Strategy-Consultant", model_name="gpt-5.4", max_loops=1),
    ],
    director_model_name="gpt-5.4",
    director_feedback_on=False,
    parallel_execution=True,   # all 3 agents run at the same time
)

result = swarm.run(task="Analyze AI infrastructure investment trends for 2025")
print(result)

Agent-as-Judge (`agent_as_judge`)¶

With agent_as_judge=True, after all worker agents complete their tasks in a cycle, a dedicated judge agent is instantiated to score every worker's output using a structured JudgeReport schema. This replaces the freeform director_feedback_on feedback — if both are True, agent_as_judge takes priority.

Scoring dimensions (each 0–10, weighted composite):

Dimension	Weight	What it measures
Task Adherence	25%	Did the agent do what it was assigned?
Accuracy & Factual Integrity	25%	Are claims correct and well-supported?
Depth & Completeness	20%	Is the response thorough or surface-level?
Clarity & Communication	15%	Is the output well-structured and actionable?
Contribution to Swarm Goal	15%	Does the output advance the collective mission?

Output schema:

class AgentScore(BaseModel):
    agent_name: str
    score: int          # 0–10 composite
    reasoning: str      # evidence-cited explanation
    suggestions: str    # concrete, actionable improvements

class JudgeReport(BaseModel):
    overall_quality: int        # 0–10 collective score
    scores: List[AgentScore]    # one entry per worker agent
    summary: str                # names best/worst agent, flags gaps

The JudgeReport is added to self.conversation under the role "JudgeAgent" and logged. With max_loops > 1, the director sees the scores on the next loop and can adjust its plan accordingly.

from swarms import Agent, HierarchicalSwarm

swarm = HierarchicalSwarm(
    name="Judged-Swarm",
    agents=[
        Agent(agent_name="Research-Analyst", model_name="gpt-5.4", max_loops=1),
        Agent(agent_name="Data-Analyst", model_name="gpt-5.4", max_loops=1),
        Agent(agent_name="Strategy-Consultant", model_name="gpt-5.4", max_loops=1),
    ],
    director_model_name="gpt-5.4",
    director_feedback_on=False,
    agent_as_judge=True,                    # enable structured scoring
    judge_agent_model_name="gpt-5.4",   # model for the judge
)

result = swarm.run(task="Analyze AI infrastructure investment trends for 2025")
print(result)

Difference from director_feedback_on:

	`director_feedback_on`	`agent_as_judge`
Output format	Unstructured string	Structured `JudgeReport` with typed fields
Numeric scores	No	Yes — per-agent and overall (0–10)
Machine-readable	No	Yes
Multi-loop utility	Director reads its own text feedback	Director reads structured scores and suggestions

Async Entry Point (`arun`)¶

arun() is an async wrapper around run() using asyncio.to_thread(). It does not change any swarm behavior — it simply prevents the synchronous run() from blocking the event loop when called from async code.

import asyncio
from swarms import Agent, HierarchicalSwarm

swarm = HierarchicalSwarm(
    name="Async-Swarm",
    agents=[
        Agent(agent_name="Researcher", model_name="gpt-5.4", max_loops=1),
        Agent(agent_name="Analyst", model_name="gpt-5.4", max_loops=1),
    ],
    director_model_name="gpt-5.4",
)

# Works inside any async context
result = asyncio.run(
    swarm.arun(task="Analyze AI infrastructure investment trends for 2025")
)
print(result)

For FastAPI or other async frameworks:

from fastapi import FastAPI
from swarms import Agent, HierarchicalSwarm

app = FastAPI()
swarm = HierarchicalSwarm(
    agents=[Agent(agent_name="Analyst", model_name="gpt-5.4", max_loops=1)],
    director_model_name="gpt-5.4",
)

@app.post("/analyze")
async def analyze(task: str):
    result = await swarm.arun(task=task)
    return {"result": result}

Advanced Usage Examples¶

All Three New Features Combined¶

import asyncio
from swarms import Agent, HierarchicalSwarm

swarm = HierarchicalSwarm(
    name="FullFeature-Swarm",
    agents=[
        Agent(agent_name="Research-Analyst", model_name="gpt-5.4", max_loops=1),
        Agent(agent_name="Data-Analyst", model_name="gpt-5.4", max_loops=1),
        Agent(agent_name="Strategy-Consultant", model_name="gpt-5.4", max_loops=1),
    ],
    director_model_name="gpt-4.1",
    director_feedback_on=False,
    parallel_execution=True,         # agents run concurrently
    agent_as_judge=True,             # structured scoring after each cycle
    judge_agent_model_name="gpt-4.1",
)

result = asyncio.run(
    swarm.arun(task="Analyze AI infrastructure investment trends and recommend the top 3 opportunities for 2025.")
)
print(result)

Financial Analysis Swarm¶

from swarms import Agent, HierarchicalSwarm

financial_analysis_swarm = HierarchicalSwarm(
    name="Financial-Analysis-Hierarchical-Swarm",
    description="A hierarchical swarm for comprehensive financial analysis",
    agents=[
        Agent(
            agent_name="Market-Research-Specialist",
            agent_description="Expert in market research, trend analysis, and competitive intelligence",
            model_name="gpt-5.4",
        ),
        Agent(
            agent_name="Financial-Analysis-Expert",
            agent_description="Specialist in financial statement analysis, valuation, and investment research",
            model_name="gpt-5.4",
        ),
    ],
    max_loops=2,
    director_model_name="gpt-5.4",
    planning_enabled=True,
    agent_as_judge=True,
)

result = financial_analysis_swarm.run(
    task="Conduct a comprehensive analysis of Tesla (TSLA) stock including market position, financial health, and investment potential"
)
print(result)

Interactive Dashboard¶

When interactive=True, the swarm renders a real-time Rich-based dashboard showing director status, agent execution, task assignments, and full output history across loops.

swarm = HierarchicalSwarm(
    name="Swarms Corporation Operations",
    agents=[research_agent, analysis_agent, strategy_agent],
    max_loops=2,
    interactive=True,
    director_model_name="gpt-5.4",
)

result = swarm.run("Conduct a research analysis on water stocks and ETFs")

Dashboard Panel	Description
Operations Status	Swarm name, director info, loop progress, runtime
Director Operations	Director's plan and current task orders
Agent Monitoring Matrix	All agents, status, tasks, and outputs across loops
Detailed View	Full output history per agent per loop

Autosave¶

autosave=True (default) saves conversation history to {workspace_dir}/swarms/HierarchicalSwarm/{swarm-name}-{timestamp}/conversation_history.json after all loops complete.

import os
os.environ["WORKSPACE_DIR"] = "my_project"  # optional — defaults to ./agent_workspace

swarm = HierarchicalSwarm(
    name="analysis-swarm",
    agents=[research_agent, financial_agent],
    autosave=True,  # default
)

Planning Feature¶

When planning_enabled=True, the director runs an initial planning pass before creating task orders — analyzing the task, developing a strategy, then distributing assignments.

	`planning_enabled=True`	`planning_enabled=False`
Best for	Complex multi-step tasks, uncertain requirements	Simple tasks, latency-sensitive workflows
Overhead	One extra LLM call per loop	None

swarm = HierarchicalSwarm(
    agents=[agent1, agent2],
    planning_enabled=True,
    director_temperature=0.7,
    director_top_p=0.9,
)

Output Types¶

Output Type	Description
`"dict-all-except-first"`	All conversation history as a dict, excluding the first message (default)
`"dict"`	Full conversation history as a dict
`"str"`	Conversation history as a plain string
`"list"`	Conversation history as a list

Streaming Callbacks¶

Pass a streaming_callback to run() or arun() to receive token-by-token output from each agent.

def streaming_callback(agent_name: str, chunk: str, is_final: bool) -> None:
    if chunk.strip():
        print(f"\r{agent_name}: {chunk}", end="", flush=True)
    if is_final:
        print(f"\n✅ {agent_name} done")

result = swarm.run(task="...", streaming_callback=streaming_callback)

Note: streaming callbacks are only invoked in sequential mode (parallel_execution=False). In parallel mode, chunks from concurrent agents would interleave unpredictably.

Best Practices¶

Best Practice	Description
Agent Specialization	Create agents with specific, well-defined expertise areas
Clear Task Descriptions	Provide detailed, actionable task descriptions
Appropriate Loop Count	Set `max_loops` based on task complexity (1–3 for most tasks)
Director Configuration	Adjust `director_temperature` (0.7–0.9) for desired creativity
Planning Strategy	Enable `planning_enabled` for complex tasks, disable for simple ones
Parallel vs Sequential	Use `parallel_execution=True` for independent agents; sequential when agents build on each other
Judge vs Feedback Director	Use `agent_as_judge=True` when you need structured, numeric scores; use `director_feedback_on=True` for freeform guidance
Multi-loop with Judge	Combine `agent_as_judge=True` with `max_loops > 1` so the director can act on scores in subsequent loops
Async Integration	Use `arun()` in FastAPI or any async framework to avoid blocking the event loop
Interactive Dashboard	Use `interactive=True` during development for real-time monitoring
Autosave	Keep `autosave=True` to preserve conversation history for debugging
Model Selection	Choose capable models for director (coordination) vs smaller/faster for agents

Error Handling¶

The HierarchicalSwarm includes comprehensive error handling with detailed logging. Common issues:

Issue	Solution
No agents provided	Pass at least one `Agent` in the `agents` list
`max_loops <= 0`	Set `max_loops` to a positive integer
Director not responding	Check `director_model_name` is a valid, accessible model
Judge not visible in output	Ensure `agent_as_judge=True` is set; the `JudgeReport` appears in conversation history under role `"JudgeAgent"` and is logged via `logger.info`

Performance Considerations¶

Consideration	Description
Parallel Execution	Use `parallel_execution=True` for independent agents; removes sequential bottleneck for large agent pools
Judge Overhead	`agent_as_judge=True` adds one extra LLM call per loop; use a small/fast model via `judge_agent_model_name`
Loop Optimization	Balance thoroughness vs cost with `max_loops`
Planning Overhead	`planning_enabled=True` adds one extra director call per loop
Model Selection	A smaller model for the director/judge (e.g. `gpt-4o-mini`) reduces cost without sacrificing coordination quality
Dashboard Impact	`interactive=True` adds minimal overhead but is not recommended for production
Top-P Sampling	Use `director_top_p=None` to disable nucleus sampling and rely only on temperature

HierarchicalSwarm¶

Key Features¶

Constructor¶

HierarchicalSwarm.__init__()¶

Parameters¶

Returns¶

Raises¶

Core Methods¶

run()¶

Parameters¶

Returns¶

Example¶

arun()¶

Parameters¶

Returns¶

Example¶

batched_run()¶

Parameters¶

Returns¶

Example¶

display_hierarchy()¶

Example¶

New Optional Features¶

Parallel Execution (parallel_execution)¶

Agent-as-Judge (agent_as_judge)¶

Async Entry Point (arun)¶

Advanced Usage Examples¶

All Three New Features Combined¶

Financial Analysis Swarm¶

Interactive Dashboard¶

Autosave¶

Planning Feature¶

Output Types¶

Streaming Callbacks¶

Best Practices¶

Error Handling¶

Performance Considerations¶

`HierarchicalSwarm`¶

`HierarchicalSwarm.init()`¶

`run()`¶

`arun()`¶

`batched_run()`¶

`display_hierarchy()`¶

Parallel Execution (`parallel_execution`)¶

Agent-as-Judge (`agent_as_judge`)¶

Async Entry Point (`arun`)¶