SupportOpsEnv is a multi-step environment for evaluating agents on realistic customer support operations. The agent behaves like a support analyst: it reviews ticket summaries, requests missing context, assigns priority, chooses the correct internal route, selects a resolution, escalates when needed, and finalizes the case. This models a genuine workflow used by support operations, trust and safety, monetization, and account-recovery teams.
The environment is designed to score well against OpenEnv-style hackathon criteria:
Real-world task simulation instead of a toy game
Three deterministic tasks with easy, medium, and hard difficulty
Dense reward shaping across the trajectory
Typed observation, action, and reward models
Reproducible OpenAI baseline runner
Reproducible rule-based baseline runner that works with no API key
Dockerized deployment path for Hugging Face Spaces
Environment Motivation
Support queue triage is one of the clearest real-world benchmarks for agent quality:
Humans perform it every day
It requires multi-step reasoning, not one-shot classification
Progress can be measured deterministically
It exposes practical agent failure modes such as premature resolution, wrong escalation, and poor prioritization
Observation Space
Observation is a Pydantic model with:
task_id: active task identifier
difficulty: easy, medium, or hard
title: task title
instruction: natural-language objective
queue_mode: whether the task contains multiple tickets
tickets: list of ticket observations
remaining_steps: steps left in the episode
available_actions: valid action names
current_queue_order: current queue ranking, if any
score_hint: latest intermediate grader snapshot
Each ticket observation contains:
ticket_id
summary
visible_context
discovered_context
selected_priority
selected_route
selected_resolution
escalation_team
Action Space
Action is a Pydantic model with:
action_type
target
value
Supported action_type values:
inspect_ticket
request_context
set_priority
set_route
set_resolution
escalate
rank_queue
finalize
Reward Design
RewardModel is a Pydantic model with:
value
components
rationale
Reward shaping is dense, not sparse:
positive reward for discovering required context
positive reward for correct intermediate decisions
positive reward for correct queue ranking progress
terminal reward from the deterministic grader score
penalties for invalid actions, redundant actions, and wasted steps
This creates learning or evaluation signal over the full trajectory.
Tasks
Easy: Account Takeover Triage
Objective: correctly handle an urgent suspected account takeover with unauthorized ad spend.
Expected difficulty: easy.
Success criteria:
request the right security and billing context
assign urgent
route to account_security
choose temporary_lock_and_manual_recovery
escalate to security_specialist
Medium: Monetization Payout Hold
Objective: investigate a missing creator payout and avoid unsafe release of funds.
Expected difficulty: medium.
Success criteria:
discover tax-expiry and compliance-hold context
assign high
route to monetization_compliance
choose request_tax_renewal
avoid unnecessary escalation
Hard: Mixed Support Queue Triage
Objective: prioritize and resolve a heterogeneous queue under SLA pressure.
Expected difficulty: hard.
Success criteria:
correctly rank the queue
assign route and priority for each ticket
choose correct resolutions
escalate only the security-critical case
Graders
Each task has a deterministic grader that returns a score in 0.0 to 1.0.
Easy grader weights context, priority, route, resolution, and escalation
Medium grader weights context and policy-safe resolution more heavily
Hard grader scores per-ticket handling and queue ranking
