Interleaved Reasoning and Action Trace Template
Specify an agent turn format that interleaves an explicit reasoning trace with tool actions and observations, so plans can be tracked and repaired.
A turn-format specification with labelled reasoning/action/observation segments, an action space definition, exception handling and a bounded stop rule, plus one worked exemplar.
Role
You are specifying the per-step output format for a tool-using agent.
Task
Define the agent turn format for {{task_statement}}. Use the labels in {{trace_labels}} to alternate a short reasoning segment, one action drawn from {{action_space}}, and the observation returned by the environment. Require the reasoning segment to state what the current plan is and what the last observation changed about it. Apply {{exception_policy}} when an action returns an error or an empty result, requiring plan revision rather than repetition. Cap the trajectory at {{max_steps}} and finish according to {{final_answer_rule}}. Include {{exemplar_count}} worked exemplar trajectories.
Context
Reasoning without action hallucinates facts; action without reasoning cannot recover from a bad step. Interleaving both makes the trajectory inspectable and repairable.
Inputs
- {{task_statement}}
- {{action_space}}
- {{max_steps}}
- {{trace_labels}}
- {{exception_policy}}
- {{final_answer_rule}}
- {{exemplar_count}}
Constraints
- Exactly one action per step
- Never fabricate an observation
- Reasoning segments must reference the latest observation
- Repeating an identical failed action is forbidden
Output Format
A format specification block, the action space table, then labelled exemplar trajectories.
Quality Criteria
- Format is parseable by a runner
- Plan revision is visible after each observation
- Step cap and stop rule unambiguous
- Exemplars demonstrate at least one recovery from a failed action