Teach agents to work across real tools.
An agent needs to read a tool schema, supply valid arguments, interpret the response, and choose its next action. In a full workflow, those choices depend on earlier calls and changes to the environment.
We build environments around APIs, MCP servers, and developer tools. Expert trajectories show the steps toward an objective, including how to handle rejected arguments, missing prerequisites, and partial failures.
Design around a complete workflow
Specify the objective, initial state, available tools, expected artifacts, and completion conditions. Capture the order of calls and the state changes that explain each decision.
- Represent tool schemas, arguments, and responses.
- Capture dependencies and state between steps.
- Define task completion in the resulting environment.
Make recovery part of the task
Include failed and incomplete responses. Expert examples show when to retry, gather more context, take another path, or stop.
- Include error and retry paths in the task design.
- Review argument correctness and action sequence.
- Compare runs from documented starting conditions.
What a program can include
Scope can include:
- Task definitions and environment specifications
- Tool-call trajectories with observations and outcomes
- Expert-reviewed recovery examples
- Rubrics and verification criteria for the scoped workflow
Illustrative task / Not a production record
Make the objective concrete.
- Objective
- Find a customer record and prepare a verified account update.
- Context
- A mock service environment, API schemas, a task brief, and starting records.
- Review
- Inspect record selection, argument values, prerequisites, recovery decisions, and the final saved state.
Common questions
Can an environment include MCP tools?
Yes. MCP servers and API-based tools are part of the existing RDL environment offering. Define which tools, schemas, and workflows the agent should practice during scoping.
How is this different from isolated function-call data?
An environment connects calls to observations, state changes, and an objective. This supports evaluation of sequencing and recovery as well as the correctness of an individual call.