Trajectory

What it is

When an agent interacts with an environment, it accumulates a trace of every prompt sent, tool output received, and internal decision made. This structure typically looks like a list of state-action-reward tuples. In production LLM agents, a trajectory can easily span dozens of steps and thousands of tokens. This data is the primary artifact used to debug why an agent went off track or to train a model to behave better in future iterations.

Why it matters

Ignoring trajectories makes it impossible to perform effective root cause analysis on agent failures. If your agent is getting stuck in an infinite loop or calling the wrong tool, the trajectory is the log you inspect to find the exact step where the reasoning diverged. Failing to store these records prevents you from performing offline evaluation or using that data to improve your agent through fine-tuning.

In practice

You capture this by implementing structured logging for every turn in your agent orchestration loop. You should log the initial state, the agent's thought process, the tool call parameters, and the subsequent environment response. Tools like LangSmith or custom event emitters in your workflow help serialize this path for later analysis.

The tradeoff

The primary tradeoff is between observability and storage costs, as long, complex trajectories consume significant log storage and make debugging harder by adding noise.

Where it appears

Research summaries that use Trajectory, each linked to its source paper.

• Selecting Better Training Data for Agents
  SWE-Prime: Fewer Trajectories, Better Performance

  The researchers introduced SWE-Prime, a method that selects a small, high-quality subset of training trajectories to improve software engineering agent performance.

• Improving AI Agent Planning for Development
  TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning Development

  The paper introduces a planning framework for AI coding agents that aligns their development processes with human practices to improve task performance.

• Improving Software Agents With Self-Distilled Knowledge
  SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution

  SkillForge enhances software engineering agents by distilling repository-specific knowledge into reusable skills to solve project-specific issues.

• Distilling Hierarchical Memory for Agent Models
  Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory

  The researchers developed a training-free framework that distills complex strategies and tool-calling logic from large teacher models into smaller, more efficient student models.

• Decoupling Agent Scaffolding to Improve Planning
  DCAS: Decoupling CLI Agent Scaffolding to Internalize Planning across Scaffolds

  The researchers introduced DCAS, an interception layer that decouples agent planning from specific CLI environments to enable cross-environment training and better performance.

• Optimizing Model Handoffs in Coding Agents
  The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents

  The paper investigates the efficiency and accuracy trade-offs when transitioning tasks between low-cost and high-cost language models in agentic coding workflows.

• Improving Visual Tool Use for Models
  OpenVisTool: An Open Recipe for Synthesizing Instructive Visual Tool-Use Trajectories

  OpenVisTool introduces a training method that teaches models to use external visual tools only when necessary, improving performance over fixed image encoding.

• Learning Human-Like Handwriting for Robots
  Robot Learning from Human Demonstrations: Handwritten Alphabet Trajectories and Human-Likeness Evaluation

  The paper presents a method for robots to replicate human handwriting by learning trajectories from human demonstrations using Gaussian models.

• Xiaomi-Robotics-1: Scaling Robot Action Models
  Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

  Xiaomi-Robotics-1 introduces a large-scale vision-language-action model and a scalable auto-labeling pipeline to overcome data bottlenecks in robotics, achieving strong performance on robot benchmarks.

• Standardizing Data Generation for AI Agents
  What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents

  The ACE framework establishes a formal structure for evaluating and improving the data generated to train autonomous AI agents.

• Improving Multi-Turn AI Agent Credit Assignment
  IAPO: Influence-Aware Policy Optimization for Credit Assignment in Multi-Turn Service Agents

  The paper introduces IAPO, a method that improves agent training by redistributing reward credit based on how agent actions influence one another within multi-turn service workflows.

• Gaussian Guidance for Better Agent Training
  Agent-G$^2$: Gaussian Guidance for Agentic Reinforcement Learning

  Agent-G2 optimizes agent training by using Gaussian-based guidance to sample expert trajectory lengths, achieving higher success rates at a fraction of the cost of traditional probing methods.

• Efficient Reinforcement Learning for Video Models
  Annotations as Rollouts: Efficient and Scalable Reinforcement Learning for Video MLLMs

  The researchers introduced a method called OraRL that integrates ground truth annotations as oracle rollouts to improve video model performance and reduce inference latency.

• Improving Reinforcement Learning for Coding Agents
  LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents

  Lego-RL is a framework that aligns native coding execution harnesses with policy-gradient training to improve agent performance and stability.

• Optimizing Reinforcement Learning for Autonomous Agents
  Agent Lightning v1.0: Towards Harnessed Agentic RL

  Agent Lightning v1.0 provides a declarative framework to manage the complex training loops required for agents that interact with external environments.

• Making AI Models Learn Their Own Context
  Latent On-Policy Self-Distillation

  The researchers developed a method to replace hand-designed improvement rules with a system that learns to generate its own contextual guidance for model training.

• Optimizing Mixed RL Rollout Throughput
  Scheduling Mixed RL Rollouts Beyond Prefix Locality

  The paper introduces a scheduling method that increases rollout throughput and reduces iteration time by managing how heterogeneous reinforcement learning workloads share KV-cache capacity.

• Improving Agent Feedback with Trajectory Distillation
  Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learning

  The paper introduces a method to improve agent performance by dynamically allocating hindsight feedback across individual decision steps in a multi-turn task.

• Improving Mobile GUI Agents with Hindsight
  The Next Screenshot Knows: Gated Hindsight Distillation for Mobile GUI Agents

  The researchers introduce Gated Hindsight Distillation to help GUI agents learn from future screenshots when current screen data is insufficient for decision making.

• Multi-tenant cloud service for VLA training
  JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models

  JoyNexus is a service-oriented framework that enables multi-tenant post-training for Vision-Language-Action models by decoupling compute resources from model execution.

• Recovering Pruned LLM Generation Capabilities
  ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

  ShortOPD uses a dynamic distillation strategy to fix structural collapse in pruned LLMs by adjusting training rollouts based on model output quality.

• Removing Iterative Rollout from Robot Policies
  Keep the Future, Drop the Rollout: RIFT for World Action Models

  The RIFT framework enables robot policies to predict future states in a single pass, eliminating the need for iterative video generation during inference.

• Evaluating Autonomous Scientific Agent Performance
  FrontierChallenge: Evaluating Scientific Workflow Completion

  The paper introduces FrontierChallenge, a benchmark for evaluating how well AI agents complete end-to-end scientific workflows.

• Building Verifiable Skills for Reinforcement Learning
  SkillForge: Evolving Verifiable Skills for Reinforcement Learning Agents

  SkillForge introduces a system that distills and verifies reusable skills for agents, significantly improving performance on complex tasks.

• Optimizing Model Handoffs in Coding Agents
  The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents

  The paper investigates the efficiency and accuracy trade-offs when transitioning tasks between low-cost and high-cost language models in agentic coding workflows.

• Scaling Cyber Security Skills With AI
  CyberFactory: Scaling Cyber Security Capabilities with Instances from the Wild

  The researchers developed CyberFactory, a framework that leverages existing vulnerability data to train an AI model, OpenAegis, to improve security analysis performance.

• Attributing Failure Points in Agentic RAG
  When Failures Propagate: Causal Failure Attribution in Agentic Retrieval-Augmented Generation

  The paper introduces an interventional benchmark to pinpoint the specific hop in a multi-hop agentic retrieval chain where a failure originated.

• Diagnosing Failure Patterns in Autonomous Research Agents
  How Do Agents Fail on AutoResearch: End-to-End Diagnostic Evaluation on 100 Real-World Frontier Research Tasks

  The paper introduces a diagnostic evaluation suite and a failure taxonomy to systematically identify the root causes of failure in autonomous research agents across the entire scientific lifecycle.

• Why Power Sampling Fails Reasoning Tasks
  More Correct Mass, Worse Answers: Why Power Sampling Can Fail and How to Fix It

  The paper identifies that Power Sampling causes accuracy drops by misaligning trajectory coverage, and introduces a Relative-Rank SoftSat method to improve consensus aggregation.

• Improving Visual Tool Use for Models
  OpenVisTool: An Open Recipe for Synthesizing Instructive Visual Tool-Use Trajectories

  OpenVisTool introduces a training method that teaches models to use external visual tools only when necessary, improving performance over fixed image encoding.

• Efficient LLM Reasoning via Selective Regeneration
  Selective Regenerative Decoding: Trajectory-Level Intervention for Inference-Time Reasoning

  Selective Regenerative Decoding optimizes LLM inference by identifying and regenerating only the degraded portions of a reasoning sequence rather than discarding the entire trajectory.

• Predicting Movement Paths Without Expensive Training
  Non-Parametric Spatiotemporal Trajectory Prediction via State-Conditioned Transition Sampling

  The paper introduces a non-parametric approach for multi-modal trajectory prediction that constructs a transition table from historical data to represent uncertainty at route junctions without relying on expensive GPU training or large-scale data.

• Adaptive Constraint Planning for Racing Cars
  Control-Informed Constraint Adaptation in Minimum-Time Trajectory Planning for Autonomous Racing

  The paper introduces a feedback loop between the tracking controller and trajectory planner that adjusts spatial constraints to prevent sub-optimal performance caused by model mismatches.

• Why Wrong AI Messages Still Matter
  Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages

  The paper introduces a method to identify helpful reasoning in incorrect AI model messages that improves the final output of multi-agent systems.

• Improving Driving Video Search via Motion
  TraVEL: Trajectory-Guided Video Embedding Learning for Driving-Video Retrieval

  TraVEL improves driving video retrieval by training embedding models to prioritize ego-vehicle movement patterns over static visual shortcuts.

• Improving Web Agent Performance via Co-Synthesis
  SynWeaver: Website-Prior Task and Trajectory Co-Synthesis for Web Agents

  SynWeaver improves web agent accuracy by co-synthesizing website-specific tasks and execution trajectories to overcome the lack of supervision on unseen websites.

• Automated Safety Harness Evolution for Agents
  SHE: Trajectory-driven Safety Harness Evolution for LLM Agents

  The SHE framework allows LLM agent safety systems to automatically evolve over time by analyzing failure trajectories to refine safety boundaries.

• Improving Agent Feedback with Trajectory Distillation
  Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learning

  The paper introduces a method to improve agent performance by dynamically allocating hindsight feedback across individual decision steps in a multi-turn task.