| environment: |
| name: customer-support-agent |
| version: "2.0.0" |
| description: > |
| OpenEnv-compliant autonomous customer support environment. |
| Each episode is a 3-phase Markov Decision Process: Triage -> Route -> Resolve. |
| An AI agent classifies, assigns, and resolves realistic BPO/customer-support |
| tickets across e-commerce and SaaS workflows. |
| Deterministic per-phase reward shaping; observation state evolves each step. |
| tags: |
| - openenv |
| - customer-support |
| - reinforcement-learning |
| - multi-step |
| - bpo |
|
|
| tasks: |
| - name: easy |
| description: > |
| 3-phase MDP with straightforward tickets. Phase 1: classify issue type. |
| Phase 2: route to correct team. Phase 3: respond or refund. |
| difficulty: easy |
| max_steps: 5 |
| num_scenarios: 10 |
| grader: openenv.graders.grader.GraderEngine |
|
|
| - name: medium |
| description: > |
| 3-phase MDP with moderate complexity. Phase 1: classify. Phase 2: route. |
| Phase 3: respond, refund, or escalate safety/product defect issues. |
| difficulty: medium |
| max_steps: 5 |
| num_scenarios: 10 |
| grader: openenv.graders.grader.GraderEngine |
|
|
| - name: hard |
| description: > |
| 3-phase MDP with ambiguous, high-stakes tickets. Requires distinguishing |
| escalate vs refund vs respond with asymmetric business penalties. |
| difficulty: hard |
| max_steps: 5 |
| num_scenarios: 10 |
| grader: openenv.graders.grader.GraderEngine |
|
|
| observation_space: |
| type: object |
| fields: |
| ticket_id: |
| type: string |
| description: "Unique ticket identifier (e.g. TKT-E001)" |
| issue_type: |
| type: string |
| description: "Category of the customer issue. 'unknown' in Phase 1; revealed after classify." |
| enum: [unknown, shipping, billing, technical, returns, safety, cancellation, complaint, orders, sales] |
| sentiment: |
| type: string |
| description: "Detected customer sentiment" |
| enum: [positive, neutral, negative, angry] |
| priority: |
| type: string |
| description: "Ticket urgency level" |
| enum: [low, medium, high, critical] |
| message: |
| type: string |
| description: "The customer's original message text" |
| history: |
| type: array |
| items: |
| type: string |
| description: "Phase transition messages and prior agent actions in this episode" |
|
|
| action_space: |
| type: object |
| fields: |
| action_type: |
| type: string |
| description: "Primary action to execute" |
| enum: [classify, assign, respond, refund, escalate] |
| team: |
| type: string |
| nullable: true |
| description: "Target team (used with assign/escalate actions)" |
| enum: [logistics_team, tech_support_team, safety_team, finance_team, orders_team, management_team] |
| response: |
| type: string |
| nullable: true |
| description: > |
| In Phase 1: the identified issue_type (e.g. 'billing'). |
| In Phase 3: customer-facing response message. |
| |
| reward: |
| range: [0.002, 0.998] |
| phases: |
| phase_1_triage: |
| classify_with_correct_issue_type: 0.998 |
| classify_only: 0.55 |
| wrong_action: 0.002 |
| phase_2_route: |
| correct_team: 0.998 |
| wrong_team: 0.400 |
| wrong_action: 0.002 |
| phase_3_resolve: |
| description: "Proportional quality score: action_type (0.3) + team (0.3) + keyword coverage (0.4)" |
| range: [0.002, 0.998] |
| penalties: |
| unnecessary_refund: -0.50 |
| unnecessary_escalation: -0.30 |
| wrong_team: -0.15 |
| excessive_steps: -0.10 per step beyond 3 |
| episode_score: "mean(phase_1_reward, phase_2_reward, phase_3_reward) clamped to [0.002, 0.998]" |
