
The Decisions That Shape Better AI Agents
New research reveals how selecting stronger agent trajectories and learning from critical decisions can improve the way AI models work.

Two AI coding agents were given the same software bug to fix. Both succeeded. One reached the solution in 36 steps, progressing through inspection, diagnosis, modification, and review. The other needed 73 steps, repeatedly interrupting processes and taking a less focused route to the same result.
The comparison comes from an October 2026 study by researchers at Peking University and Huawei. It illustrates an increasingly important consideration in AI training: when a model learns from another agent's work, the sequence of actions can be as valuable as the completed result.
Agent trajectories capture those sequences. They record the tool calls, observations, decisions, intermediate results, and actions taken as a model works through a task. These records can be used to train other models, providing examples of how to investigate problems, navigate software environments, recover from errors, and complete multi-step work.
As developers collect larger volumes of agent experience, a new opportunity is emerging: identifying which behaviors are most useful to learn, and which decisions make the greatest difference to success.
Recent research is making both questions more tractable.
Finding the strongest examples in successful work
Supervised fine-tuning (SFT) trains models using demonstrations of desired behavior. For an AI agent, these demonstrations can include entire interactions with an environment rather than a single prompt and response.
A coding trajectory might contain dozens of steps: examining a repository, identifying relevant files, testing hypotheses, modifying code, running tests, and responding to errors. Training on the trajectory exposes the model to the actions and observations involved in reaching the solution.
The quality of those interactions matters.
In their paper, Quality, Not Just Outcome, the researchers examined 155,144 successful coding-agent trajectories and developed a method for assessing the structure of each run.
Their approach, called Trajectory Quality Score (TQS), measures characteristics associated with coherent problem-solving. These include how an agent narrows its attention toward relevant information, whether its actions remain aligned with its stated intentions, and how its behavior evolves as it approaches a solution.
The researchers then tested whether selecting trajectories with stronger structural characteristics would improve subsequent model training.
They created three training sets containing 1,024 successful trajectories each: one selected using high TQS scores, one selected randomly, and one containing lower-ranked trajectories. All three were used to fine-tune the same coding model under identical training conditions.
The resulting models were evaluated on SWE-bench Verified, a benchmark of real software engineering problems.
The differences were substantial.
The model trained on highly ranked trajectories resolved 27.6% of benchmark tasks, compared with 24.8% for randomly selected trajectories and 21.8% for the lowest-ranked examples. The original model, without this additional fine-tuning, scored 24.4%.
The researchers also compared their method against simpler selection strategies. Choosing the shortest trajectories did not deliver the same improvement, while a more expensive LLM-based selection method performed slightly better, reaching 28.4%.
The experiment was limited to one model configuration and a fixed training budget, without repeated training runs to estimate variability. Still, it provides evidence that the structure of successful agent behavior can influence what another model learns from it.
This opens up a useful direction for training-data development. Instead of relying exclusively on successful outcomes to select demonstrations, researchers can examine the behaviors that produced those outcomes.
The objective is not to make every agent follow the shortest possible route. Exploration, verification, and recovery can be essential parts of difficult work. The more valuable distinction is between activity that advances the task and activity that adds little useful information.
Learning from the decisions that change an outcome
Selecting strong demonstrations is one way to improve training. Another is to examine unsuccessful attempts and identify where a different decision could have changed the result.
This is the focus of Verified Critical Step Optimization for LLM Agents, published in ACL Findings 2026.
The researchers approached agent training through a different unit of supervision: individual decision points within longer trajectories.
Their method begins with failed attempts. A process reward model identifies candidate steps that may have contributed to failure, and an expert model proposes alternative actions. The original agent is then allowed to continue from those alternatives.
If the agent successfully completes the task, the alternative becomes a verified example of a decision that can lead to a better outcome.
These verified comparisons are used for preference optimization, helping the model learn which actions to favor when it encounters similar circumstances.
The method concentrates feedback on decisions with demonstrated consequences.
Across two research-agent benchmarks, the researchers reported relative improvements of 37% and 26% over their supervised fine-tuning baselines, while providing supervision at approximately 16% of trajectory steps.
Those results are specific to the experimental setup. Identifying critical steps remains difficult, and success after a particular intervention does not establish that the same choice will work in every context.
Nevertheless, the method suggests a productive approach to agent training. Rather than treating every action in a long interaction as equally informative, researchers can investigate where targeted feedback is most useful.
There is an important connection between these two studies.
Trajectory selection identifies examples of behavior worth reproducing. Critical-step optimization identifies decisions worth improving.
One operates across an entire successful interaction; the other concentrates on consequential moments within failed attempts.
Together, they demonstrate how agent experience can provide increasingly specific training signals.
Making progress visible throughout a task
These training approaches also raise questions about evaluation.
A final task outcome provides essential evidence of performance. For many agent workflows, the result can be checked directly: a code change passes tests, a database reaches the requested state, or a software operation completes successfully.
But longer tasks contain meaningful progress before the final result is reached.
An agent may correctly identify a problem, retrieve the necessary information, and complete several required operations before encountering a particular obstacle. Another agent may fail much earlier. Understanding that difference can help researchers identify where additional training is needed.
In Beyond Final Success, published at EMNLP 2026, researchers developed a framework for analyzing progress through verifiable task milestones.
They applied it to 4,740 agent trajectories from 22 models across 108 tasks, examining which milestones were reached, when they were reached, and how progress related to resource use.
The analysis revealed differences between models with identical final success rates, including how effectively they progressed through intermediate requirements.
This provides a more detailed picture of agent capability.
A benchmark can report how often a model succeeds. A trajectory-based analysis can additionally reveal where its performance is strong, where progress stalls, and whether its actions are contributing to the task.
These observations can inform more targeted development. They also help distinguish genuine improvements in task execution from changes that merely affect how an agent presents its final result.
The human-data opportunity
The ability to collect, analyze, and learn from agent trajectories creates new possibilities for AI training. Automated checks can verify outcomes, structural methods can identify useful patterns, and models can help locate candidate decisions for further investigation.
Human expertise becomes especially valuable where interpreting behavior requires understanding the task and the evidence available.
An agent that revisits a document may be wasting effort, or it may be resolving an important contradiction. An agent that changes its plan may be recovering intelligently from new information. A tool call may appear successful while the underlying operation remains incomplete.
Making these distinctions reliably requires more than recognizing a sequence of actions.
It involves understanding what the agent was trying to accomplish, which information was available at each step, what its actions established, and which alternatives were legitimate under the task's requirements.
This connects trajectory-based training to broader human-data capabilities.
The evidence-handling skills used in factuality evaluation help reviewers determine whether an agent's next action follows from its observations. Preference judgment helps distinguish stronger strategies when multiple approaches are valid. Temporal and entity precision become important when tracking state changes across tools, video, audio, and other modalities.
As agent workflows become more sophisticated, these capabilities can support increasingly targeted forms of supervision.
A reviewer might evaluate an entire trajectory, identify a consequential error, verify an intermediate outcome, or assess whether a proposed correction preserves the task's constraints.
The contribution is not necessarily a larger quantity of annotations. It is more informative evidence about the behavior being learned.
For teams developing training pipelines, the opportunity is to combine scalable automated analysis with selective expert judgment. Automated systems can surface promising examples or suspicious decisions; human reviewers can help resolve ambiguity, validate consequential differences, and establish whether the behavior is appropriate for the intended use.
The balance will vary by task. Software tests can verify many coding outcomes automatically. Research, multimodal interpretation, and complex business workflows often require additional judgment about whether the result satisfies the underlying objective.
Building capability from experience
Agent trajectories offer a richer source of training data because they capture how work unfolds.
They make it possible to study the relationship between information gathering, tool use, decision-making, recovery, and successful execution. Researchers can select demonstrations with useful behavioral patterns, investigate decisions that change outcomes, and evaluate progress throughout longer tasks.
The recent studies are early examples of this potential. Their results are encouraging, but broader validation is needed across models, environments, and task types.
What is becoming clearer is the role of data quality in turning agent experience into improved capability.
The value of a trajectory lies in the behaviors it can teach. The value of feedback lies in the distinctions it helps a model learn. And the value of evaluation lies in establishing whether those lessons translate into better performance on new work.
As AI agents take on more demanding responsibilities, developing those connections will become an increasingly important part of model training.
The next generation of capable agents will benefit from learning experiences that capture how good work is done - and how difficult decisions can be made better.
Data Gradient develops practical training for people contributing to AI model training and evaluation. Through Data Gradient Learn, contributors build the skills needed to interpret model behavior, evaluate evidence, and produce reliable human data across increasingly complex AI workflows.