====== Synthetic Data Generation Agents ======

**Synthetic Data Generation Agents** are agentic AI pipelines that autonomously create high-quality training datasets by decomposing complex data generation tasks into manageable subtasks executed by specialized LLM-based agents. The **AgentSynth** framework, published as a conference paper at **ICLR 2026**, demonstrates this approach for generating diverse computer-use task trajectories at scale.

===== Overview =====

Training capable AI agents requires large volumes of high-quality, diverse task data with corresponding trajectories. Human annotation is expensive (often hundreds of dollars per trajectory) and difficult to scale. Agentic synthetic data generation addresses this by leveraging **information asymmetry** — the principle that executing a task step-by-step is significantly easier than reasoning about the complete solution at once.

By decomposing generation into forward-execution subtasks, agentic pipelines produce datasets that are simple to create but challenging to solve, providing both training data and discriminative benchmarks.

===== AgentSynth Framework =====

AgentSynth is a scalable, cost-efficient pipeline for automatically synthesizing task and trajectory datasets for generalist computer-use agents. Developed at UC Berkeley by Jingxu Xie, Dylan Xu, Xuandong Zhao, and Dawn Song.

=== Multi-Agent Architecture ===

AgentSynth deploys six distinct LLM-based agents in a coordinated pipeline:

  * **Task Proposer** — Generates candidate task descriptions based on available tools and environments
  * **Task Executor** — Attempts to execute proposed tasks, producing action trajectories
  * **Task Verifier** — Validates that executed trajectories correctly complete the proposed task
  * **Task Reviser** — Refines failed or ambiguous tasks based on execution feedback
  * **Follow-up Task Proposer** — Generates compositional follow-up tasks that build on completed subtasks
  * **Task Summarizer** — Produces clean task descriptions and metadata for the final dataset

=== Difficulty Modulation ===

A key innovation is precise control over task complexity by varying the number of composed subtasks. Each individual subtask is straightforward, but chaining them creates increasingly challenging long-horizon tasks:

^ Difficulty Level ^ Subtasks ^ Agent Success Rate ^
| Level 1 | 1 | 18% |
| Level 2 | 2 | 12% |
| Level 3 | 3 | 8% |
| Level 6 | 6 | 4% |

This steep performance degradation demonstrates the benchmark's discriminative power and highlights substantial room for agent improvement.

===== Code Example =====

Simplified agentic synthetic data generation pipeline:

<code python>
from dataclasses import dataclass

@dataclass
class SubTask:
    description: str
    tools_required: list[str]
    trajectory: list[dict]
    verified: bool = False

class AgentSynthPipeline:
    def __init__(self, llm_client, environment):
        self.llm = llm_client
        self.env = environment

    def propose_subtask(self, context: dict) -> SubTask:
        prompt = f"Propose a simple computer task using: {context['available_tools']}"
        response = self.llm.generate(prompt)
        return SubTask(
            description=response.text,
            tools_required=response.tools,
            trajectory=[]
        )

    def execute_subtask(self, subtask: SubTask) -> SubTask:
        trajectory = []
        state = self.env.reset()
        for step in range(self.env.max_steps):
            action = self.llm.generate(
                f"Task: {subtask.description}\nState: {state}\nNext action:"
            )
            state, done = self.env.step(action)
            trajectory.append({"state": state, "action": action})
            if done:
                break
        subtask.trajectory = trajectory
        return subtask

    def verify_subtask(self, subtask: SubTask) -> bool:
        verification = self.llm.generate(
            f"Did this trajectory complete the task?\n"
            f"Task: {subtask.description}\n"
            f"Trajectory: {subtask.trajectory}"
        )
        subtask.verified = verification.text.lower().startswith("yes")
        return subtask.verified

    def compose_tasks(self, subtasks: list[SubTask], difficulty: int) -> dict:
        selected = subtasks[:difficulty]
        composed_description = " Then, ".join(s.description for s in selected)
        composed_trajectory = []
        for s in selected:
            composed_trajectory.extend(s.trajectory)
        return {
            "description": composed_description,
            "difficulty": difficulty,
            "trajectory": composed_trajectory,
            "num_subtasks": len(selected)
        }
</code>

===== Cost Efficiency =====

AgentSynth achieves an average cost of **$0.60 per trajectory**, orders of magnitude cheaper than human annotations. Over 6,000 diverse and realistic tasks were generated using the pipeline, integrated with the OSWorld environment for authentic computer tool interactions.

===== Broader Ecosystem =====

Other frameworks complement AgentSynth in the synthetic data generation space:

  * **NVIDIA NeMo** — Infrastructure for configuring seed datasets, column structures, and LLM-prompted generation with quality evaluation
  * **Tonic Fabricate** — Conversational agentic interface for natural-language dataset specification with real-time generation
  * **Schema-Aware Generation** — Maintaining referential integrity across related tables while generating statistically consistent synthetic records

===== References =====

  * [[https://arxiv.org/abs/2506.14205|arXiv:2506.14205 — AgentSynth: Scalable Task Generation for Generalist Computer-Use Agents]]
  * [[https://openreview.net/forum?id=CoBxmXThM6|OpenReview — AgentSynth (ICLR 2026)]]
  * [[https://github.com/sunblaze-ucb/AgentSynth|AgentSynth GitHub Repository]]
  * [[https://iclr.cc/virtual/2026/poster/10010827|ICLR 2026 Poster — AgentSynth]]

===== See Also =====

  * [[self_evolving_agents|Self-Evolving Agents]]
  * [[agentic_skills|Agentic Skills]]
  * [[synthetic_data|Synthetic Data for AI Training]]
  * [[computer_use_agents|Computer Use Agents]]