Swarm Multi-Agent Orchestration

What is Swarm Mode?

Swarm mode deploys multiple persistent, autonomous agents that work in parallel to solve complex tasks. An LLM-powered Lead Agent dynamically coordinates the team — planning tasks, spawning agents, reassigning work, and making decisions based on real-time events. Unlike static orchestration where tasks are pre-decomposed and never change, the Lead Agent continuously monitors progress and adapts: creating new tasks, canceling redundant work, and reassigning idle agents as the situation evolves.

How It Works

Shannon’s swarm workflow is driven by an event-driven Lead Agent loop:

Lead Agent Event Loop

The Lead Agent wakes on specific events and decides what to do next:

Events that wake the Lead:
├── agent_idle       — An agent finished its current task or became available
├── agent_completed  — An agent produced final output
├── help_request     — An agent requested the Lead to spawn a helper
├── checkpoint       — Periodic timer (every 120s) to review overall progress
└── human_input      — User sent new instructions mid-execution

On each wake, the Lead receives the full team status (agents, tasks, budget) and chooses one or more actions.

Lifecycle Overview

Initial Planning — The Lead receives the user query and creates an initial set of tasks with optional dependency chains
Agent Spawning — The Lead spawns agents and assigns tasks, respecting dependency order
Event-Driven Coordination — As agents complete work, report idle, or hit checkpoints, the Lead dynamically reassigns tasks, revises the plan, or spawns new agents
Synthesis — When all tasks are complete, the Lead can spawn a dedicated synthesis agent or declare done to produce the final response

Task Dependencies (DAG)

Tasks can declare dependencies on other tasks, forming a directed acyclic graph (DAG):

Tasks:
├── task-1: "Research US AI chip market"        (depends_on: [])
├── task-2: "Research Japan AI chip market"     (depends_on: [])
├── task-3: "Research South Korea AI chip market" (depends_on: [])
└── task-4: "Write comparative analysis"        (depends_on: [task-1, task-2, task-3])

The system enforces dependency order — task-4 cannot be assigned until all three research tasks are complete. The Lead can dynamically create new dependency chains via revise_plan.

Lead Agent Actions

Each time the Lead wakes, it selects one or more actions:

Action	Description
`spawn_agent`	Create a new agent for a specific task
`assign_task`	Assign a pending task to an idle agent
`revise_plan`	Dynamically create new tasks or cancel existing ones
`send_message`	Send a message to a specific agent
`broadcast`	Send a message to all agents
`file_read`	Read a workspace file (zero LLM cost, max 3 rounds)
`shutdown_agent`	Terminate a specific agent
`interim_reply`	Push a progress update to the user
`noop`	Do nothing (no action needed right now)
`done`	Declare all work complete, proceed to closing phase
`reply`	Return the final response directly to the user (closing phase only)
`synthesize`	Trigger the synthesis pipeline instead of replying directly

Agent Actions

Each iteration, an agent chooses exactly one action:

Action	Description
`tool_call`	Execute a tool (web search, file read, etc.)
`publish_data`	Share findings with the team via the workspace
`send_message`	Send a direct message to a specific teammate
`request_help`	Ask the Lead to spawn a new helper agent
`idle`	Signal that the current task is complete and await reassignment
`done`	Return final response (auto-converts to `idle`)

Agents cannot self-exit. When an agent returns done, it automatically converts to idle status. Only the Lead Agent can terminate agents via shutdown_agent. This ensures the Lead maintains full control over team composition.

Inter-Agent Communication

Swarm agents collaborate through two mechanisms:

P2P Messaging

Agents send direct messages to specific teammates through Redis-backed mailboxes. Message types include request, offer, accept, delegation, and info. Before each LLM call, the agent’s mailbox is checked for new messages. Incoming messages appear in the agent’s prompt context.

Shared Workspace

Agents publish findings to topic-based workspace lists. Before each iteration, every agent fetches recent workspace entries from all topics, so the entire team stays aware of collective progress.

Shared Workspace:
├── Topic: "findings"
│   ├── Agent-Takao: "NVIDIA dominates US with 80% market share..."
│   └── Agent-Mitaka: "Japan focuses on edge AI chips..."
└── Topic: "sources"
    └── Agent-Kichijoji: "Samsung foundry plans announced..."

Knowledge Deduplication

Shannon prevents redundant work across agents with three layers of deduplication:

L1: Per-Agent URL Fetch Cache

Each agent caches URLs it has already fetched. If the same URL is requested again within the same agent loop, the cached content is returned without a network call.

L2: Cross-Agent Shared URL Metadata

URL metadata (title, summary, key facts) is shared across all agents in the team. When Agent B tries to fetch a URL that Agent A already processed, it receives the cached metadata instead of re-fetching — saving both time and tokens.

L3: Search Overlap Detection

URLs discovered by search results are tracked across all agents. When a new search returns URLs where 70% or more have already been discovered by other agents, the system injects a warning to find new angles. Additionally, a search saturation detector compares recent queries using Jaccard word-level similarity (threshold 0.7, window of 3 queries) to flag repetitive searches.

Convergence Detection

Three mechanisms prevent agents from running indefinitely:

No-Progress Detection

If an agent takes 3 consecutive iterations with no meaningful action (empty or unrecognized actions), it is considered converged and transitions to idle status. Note that tool_call, send_message, and publish_data all reset this counter.

Consecutive Error Abort

If 3 consecutive permanent tool errors occur (not transient errors like rate limits), the agent aborts and reports the failure.

Max Iterations Force-Done

On the last iteration, if the agent has not called done or idle, the workflow forces completion and builds a summary from the most recent iterations.

Transient errors (rate limits, timeouts, 503s) trigger automatic retry with escalating backoff (5s increments, max 30s) and do not count toward the abort threshold.

Global Budget Control

Swarm execution is bounded by three budget layers that prevent runaway costs:

Budget Layer	Default	Description
`max_total_llm_calls`	`200`	Maximum LLM calls across all agents
`max_total_tokens`	`1,000,000`	Maximum tokens consumed across all agents
`max_wall_clock_minutes`	`30`	Maximum wall-clock time for the entire swarm

The Lead Agent receives budget information (remaining calls, tokens, time) in its context, enabling it to make cost-aware decisions — such as shutting down low-priority agents or skipping optional tasks when budget is tight.

When to Use Swarm vs Other Workflows

Scenario	Recommended Workflow
Simple Q&A, single-step tasks	Simple / DAG
Multi-step research with citations	Research Workflow
Multi-agent code review, testing, and fixes	Swarm
Financial analysis with multiple analyst perspectives	Swarm
Data processing pipelines with Python/Bash execution	Swarm
Tasks where agents need to share intermediate findings	Swarm
Long-running exploration with dynamic subtask discovery	Swarm
Tasks with complex dependency chains between subtasks	Swarm

Swarm mode uses more tokens than standard workflows because each agent runs multiple LLM iterations and the Lead Agent consumes tokens for coordination decisions. Use it for tasks that genuinely benefit from persistent, collaborative multi-agent execution.

Configuration

Swarm behavior is controlled via config/features.yaml:

Parameter	Default	Description
`swarm.enabled`	`true`	Enable/disable swarm workflow
`swarm.max_agents`	`10`	Maximum total agents (initial + dynamic)
`swarm.max_iterations_per_agent`	`25`	Max reason-act loops per agent
`swarm.agent_timeout_seconds`	`1800`	Per-agent timeout (30 minutes)
`swarm.max_messages_per_agent`	`20`	P2P message cap per agent
`swarm.workspace_snippet_chars`	`800`	Max chars per workspace entry in prompt
`swarm.workspace_max_entries`	`5`	Max recent entries shown to each agent
`swarm.max_total_llm_calls`	`200`	Global LLM call budget for the entire swarm
`swarm.max_total_tokens`	`1000000`	Global token budget for the entire swarm
`swarm.max_wall_clock_minutes`	`30`	Maximum wall-clock time for the swarm

Streaming Events

Swarm workflows emit SSE events for real-time monitoring:

Event Type	Agent ID	When
`WORKFLOW_STARTED`	`swarm-supervisor`	Workflow begins
`PROGRESS`	`swarm-lead` / `swarm-supervisor`	Planning, spawning, reassigning
`LEAD_DECISION`	`swarm-lead`	Lead made a planning decision (spawn, assign, revise, etc.)
`TASKLIST_UPDATED`	`swarm-lead`	Task dependency graph changed (tasks created or canceled)
`TEAM_STATUS`	`swarm-lead`	Team composition changed (agent spawned or shut down)
`AGENT_STARTED`	Agent name	Agent begins first iteration
`AGENT_COMPLETED`	Agent name	Agent finishes
`WORKFLOW_COMPLETED`	`swarm-supervisor`	Final synthesis complete

Next Steps

Swarm Tutorial

Step-by-step guide to running swarm workflows

Workflows & Patterns

Other workflow types and cognitive patterns

Streaming

Real-time event streaming

Cost Control

Budget management for multi-agent tasks

Getting Started

Core Concepts

Guides

Swarm Multi-Agent Orchestration

What is Swarm Mode?

How It Works

Lead Agent Event Loop

Lifecycle Overview

Task Dependencies (DAG)

Lead Agent Actions

Agent Actions

Inter-Agent Communication

P2P Messaging

Shared Workspace

Knowledge Deduplication

Convergence Detection

Global Budget Control

When to Use Swarm vs Other Workflows

Configuration

Streaming Events

Next Steps

Swarm Tutorial

Workflows & Patterns

Streaming

Cost Control

Getting Started

Core Concepts

Guides

​What is Swarm Mode?

​How It Works

​Lead Agent Event Loop

​Lifecycle Overview

​Task Dependencies (DAG)

​Lead Agent Actions

​Agent Actions

​Inter-Agent Communication

​P2P Messaging

​Shared Workspace

​Knowledge Deduplication

​Convergence Detection

​Global Budget Control

​When to Use Swarm vs Other Workflows

​Configuration

​Streaming Events

​Next Steps

Swarm Tutorial

Workflows & Patterns

Streaming

Cost Control

What is Swarm Mode?

How It Works

Lead Agent Event Loop

Lifecycle Overview

Task Dependencies (DAG)

Lead Agent Actions

Agent Actions

Inter-Agent Communication

P2P Messaging

Shared Workspace

Knowledge Deduplication

Convergence Detection

Global Budget Control

When to Use Swarm vs Other Workflows

Configuration

Streaming Events

Next Steps