Dynamically fork sub-agents at any step based on task requirements instead of using fixed agent counts. Linearize hierarchical trajectories into unified sequences for GRPO training. Demonstrates 6.36x more sub-agents than baselines via learned orchestration.
Dynamically fork sub-agents at any step based on task requirements instead of using fixed agent counts. Linearize hierarchical trajectories into unified sequences for GRPO training. Demonstrates 6.36x more sub-agents than baselines via learned orchestration.
WideSeek: Breadth Scaling via Dynamic Multi-Agent Forking
Traditional multi-agent systems use fixed agent counts, but task complexity varies: some steps need multiple exploratory agents, others work fine with a single path. WideSeek enables dynamic agent forking—the main agent autonomously creates variable numbers of sub-agents at each step. This is trained end-to-end via GRPO by linearizing hierarchical execution traces into single unified sequences.
The key insight is that agent forking is itself a learnable skill. By treating multi-agent trajectories as extended single sequences, WideSeek applies standard policy optimization to learn when and how many agents to spawn.
Core Concept
WideSeek operates on two principles:
Dynamic Agent Forking: Main agent can call call_subagent(task) any number of times, creating sub-agents on-demand to parallelize work.
Unified Trajectory Training: All agent trajectories (main + sub-agents) are linearized into a single sequence, then optimized with GRPO as if from one agent.
This unifies multi-agent coordination with policy optimization.
Architecture Overview
Main Agent: Orchestrates task decomposition; can fork sub-agents dynamically
Sub-Agent Pool: Execute assigned subtasks in parallel with isolated contexts
Trajectory Linearization: Convert tree of agent decisions into linear sequence
GRPO Optimizer: Standard group relative policy optimization on linearized trajectories
Shared Model: Single LLM backbone for all agents (main and sub)
Tool Restrictions: Sub-agents have subset of main agent's tools (e.g., search, no forking)
Implementation
Step 1: Design Agent Forking Interface
Create the tool that allows dynamic sub-agent creation.
"""
Create and return a new sub-agent.
Args:
subtask: Specific task for sub-agent
tools: Available tools for sub-agent
parent_context: Context from parent agent
Returns:
Sub-agent instance
"""
if
not
self
return
None
self
1
self
1
self
return
class
SubAgent
def
__init__
self, task: str, available_tools: List[str],
context: str, depth: int
"""Lightweight sub-agent for specific subtask."""
self
self
self
self
self
self
None
def
execute
self, model
str
"""Execute subtask and return result."""
# Construct prompt with restricted tools
f"""You are a sub-agent working on: {self.task}
Available tools: {', '.join(self.available_tools)}
Note: You cannot fork new agents. Complete this task and report result.
Context: {self.context}"""
0.7
512
self
"type"
"thought"
"content"
self
return
Step 2: Implement Trajectory Linearization
Convert multi-agent tree execution into linear sequence for GRPO.
# Trajectory linearizationclassTrajectoryLinearizer:
def__init__(self):
"""Convert hierarchical agent executions to linear sequences."""self.linearized = []
deflinearize_execution(self, main_trajectory: List[dict],
subagent_trajectories: Dict[int, List[dict]]) -> List[dict]:
"""
Convert tree of agent trajectories into single linear sequence.
Args:
main_trajectory: Main agent's thought/action sequence
subagent_trajectories: Dict mapping agent_id to trajectory
Returns:
Single linearized trajectory for GRPO
"""
linearized = []
# Add main agent thoughtsfor step in main_trajectory:
if step.get("type") == "thought":
linearized.append({
"agent": "main",
"content": step["content"],
"type": "thought"
})
elif step.get("type") == "fork":
# Record fork decision
fork_id = step.get("fork_id")
task = step.get("task")
linearized.append({
"agent": "main",
"content": f"[FORK {fork_id}: {task}]",
"type": "fork"
})
# Append sub-agent trajectory inlineif fork_id in subagent_trajectories:
for substep in subagent_trajectories[fork_id]:
linearized.append({
"agent": f"sub_{fork_id}",
"content": substep.get("content"),
"type": "thought"
})
# Record sub-agent result
result = step.get("result")
linearized.append({
"agent": f"sub_{fork_id}",
"content": f"[RESULT: {result}]",
"type": "result"
})
elif step.get("type") == "action":
linearized.append({
"agent": "main",
"content": step["content"],
"type": "action"
})
return linearized
defcompute_sequence_loss(self, linearized: List[dict],
logits: torch.Tensor) -> torch.Tensor:
"""
Compute loss for linearized sequence (for gradient updates).
Args:
linearized: Linearized trajectory
logits: Model logits for each token
Returns:
Loss tensor
"""# Standard language modeling loss on flattened sequence
loss = torch.nn.functional.cross_entropy(
logits.view(-1, logits.shape[-1]),
self._tokenize_trajectory(linearized).view(-1)
)
return loss
def_tokenize_trajectory(self, linearized: List[dict]) -> torch.Tensor:
"""Convert trajectory to token IDs."""
text = " ".join([step["content"] for step in linearized])
# Pseudo-code: actual implementation uses model's tokenizerreturn torch.tensor([])
Step 3: Implement GRPO Training for Multi-Agent Trajectories
Extend GRPO to optimize linearized multi-agent sequences.
# Multi-agent GRPO trainingdeftrain_multi_agent_grpo(
model: nn.Module,
dataset: List[dict],
num_epochs: int = 10,
group_size: int = 4,
max_agents: int = 8):
"""
Train model with dynamic agent forking via GRPO.
Args:
model: Shared backbone for all agents
dataset: Tasks to solve
num_epochs: Training epochs
group_size: GRPO group size
max_agents: Max sub-agents per task
"""
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-5)
for epoch inrange(num_epochs):
# Group tasks for GRPOfor group_idx inrange(0, len(dataset), group_size):
group = dataset[group_idx:group_idx + group_size]
group_rewards = []
trajectories = []
for task in group:
# Execute main agent (with dynamic forking)
main_agent = Agent(
task=task["description"],
can_fork=True,
max_agents=max_agents
)
# Main agent execution: may fork sub-agents
main_trajectory, subagent_trajectories = main_agent.execute(model)
# Linearize for GRPO
linearizer = TrajectoryLinearizer()
linearized = linearizer.linearize_execution(
main_trajectory,
subagent_trajectories
)
# Evaluate final result
final_output = main_agent.result
reward = evaluate_solution(final_output, task["target"])
group_rewards.append(reward)
trajectories.append((linearized, reward))
# GRPO: compute advantages and update
group_tensor = torch.tensor(group_rewards, dtype=torch.float32)
group_mean = group_tensor.mean()
for linearized, reward in trajectories:
advantage = reward - group_mean
if advantage != 0:
# Compute loss on linearized trajectory
logits = model.forward(linearized)
loss = linearizer.compute_sequence_loss(linearized, logits)
# Scale by advantage
scaled_loss = loss * advantage
scaled_loss.backward()
optimizer.step()
optimizer.zero_grad()
print(f"Epoch {epoch}: avg_reward={group_tensor.mean():.4f}")
return model
classAgent:
def__init__(self, task: str, can_fork: bool = True,
max_agents: int = 8):
"""Main agent with optional forking."""self.task = task
self.can_fork = can_fork
self.max_agents = max_agents
self.trajectory = []
self.subagent_trajectories = {}
self.result = Noneself.fork_count = 0defexecute(self, model) -> Tuple[List[dict], Dict]:
"""Execute agent with possible sub-agent forking."""
system_prompt = f"Solve this task: {self.task}"for step inrange(10): # Max reasoning steps# Generate thought
thought = model.generate(system_prompt, max_tokens=256)
self.trajectory.append({
"type": "thought",
"content": thought
})
# Decide if forking helps (learned by GRPO)ifself.can_fork andself.fork_count < self.max_agents:
fork_decision = should_fork(model, thought, self.task)
if fork_decision:
# Create sub-agent for specific sub-task
subtask = extract_subtask(thought)
subagent = SubAgent(
task=subtask,
available_tools=["search", "calculate"],
context=self.task,
depth=1
)
result = subagent.execute(model)
self.fork_count += 1self.subagent_trajectories[self.fork_count] = subagent.trajectory
self.trajectory.append({
"type": "fork",
"fork_id": self.fork_count,
"task": subtask,
"result": result
})
# Try to solve (check if done)if is_task_complete(thought):
self.result = extract_answer(thought)
breakreturnself.trajectory, self.subagent_trajectories
Step 4: Evaluation and Metric Computation
Track multi-agent metrics.
# Multi-agent evaluationdefevaluate_multi_agent_solution(
result: str,
target: str,
num_agents_spawned: int = 0,
total_tokens: int = 0) -> dict:
"""
Evaluate solution accounting for multi-agent aspects.
Returns:
Dict with correctness, efficiency, and agent metrics
"""
correct = is_correct(result, target)
# Reward includes solution quality
base_reward = 1.0if correct else0.0# Penalize excessive agent spawning if not needed
agent_penalty = 0.1 * num_agents_spawned if num_agents_spawned > 3else0.0# Reward efficient multi-agent useif num_agents_spawned > 0and correct:
efficiency_bonus = 0.1# Bonus for parallelizationreturn {
"reward": base_reward - agent_penalty + (efficiency_bonus if num_agents_spawned > 0else0),
"correct": correct,
"agents_spawned": num_agents_spawned,
"tokens_total": total_tokens
}
Practical Guidance
When to use WideSeek:
Complex tasks with variable decomposition needs
Scenarios where parallelization is beneficial
Problems where agent depth/breadth tradeoff varies per instance
End-to-end learning of orchestration is desired
When not to use:
Simple single-path tasks
Real-time systems where agent coordination overhead matters
Scenarios requiring consistent agent behavior
Tasks with fixed decomposition structures (use static multi-agent)
Common Pitfalls:
Too many agents: Linearization becomes unwieldy; cap at 4-8
Fork scheduling imbalance: Some tasks over-fork, others under-fork
Communication gaps: Sub-agents can't see sibling results; pass context
Training instability: Multi-agent GRPO more volatile; use smaller learning rates
Key results: 6.36× more sub-agents than baselines via learned scheduling. WideSeekBench with 5,156 tasks. Code and trained 8B models released. Demonstrates breadth scaling as alternative to depth.
