Skip to content
All products

Agent training

Agent trajectory correction.

Expert reviews of agent traces that locate errors and specify the next action needed to recover.

Discuss your use case

Turn failed runs into recovery examples.

An agent can fail by choosing the wrong record, misreading an observation, or continuing after a faulty assumption. The useful training example identifies that decision and preserves the context around it.

Experts review the trace, locate the error, and explain what the agent should inspect or do next. Your team gets corrections tied to the task and its expected outcome.

Find the decision that changed the outcome

Review the objective and available information before labeling an action. Explain the correction and why it fits the current state.

  • Locate the first consequential divergence.
  • Retain the observation and task context.
  • Separate planning, execution, and verification errors.

Connect corrections to evaluation

Group recurring errors to set training priorities. Use separate evaluation tasks to check whether the corrected behavior holds in new situations.

  • Capture corrected actions and reviewer reasoning.
  • Categorize repeated failure patterns.
  • Evaluate recovery on held-out task variants.

What a program can include

Scope can include:

  • Expert-reviewed agent traces
  • Error locations and failure categories
  • Corrected next actions with supporting context
  • Review notes to inform follow-up training and evaluation

Illustrative task / Not a production record

Make the objective concrete.

Objective
Recover after an API query returns several similar records.
Context
The task brief, candidate records, prior actions, and the latest tool response.
Review
Identify the missing disambiguation step and show how to verify the correct record before making an update.

Common questions

Is trajectory correction limited to browser agents?

No. The existing RDL offering covers agent traces across service workflows and computer use. The review needs enough observations and action context to support a defensible correction.

Does every failed run need to be rewritten?

Not necessarily. A scoped review can focus on the relevant decision point, preserving valid earlier work and documenting the correction and expected continuation.

Built around your model

What does your model need to do?

Share representative tasks and examples of a good result. We’ll scope the data and evaluation with you.

Discuss your use case