TRON

ENTERPRISE-OPERATIONS AGENTIC CAPABILITIES

Can AI agents follow enterprise policies to reliably complete multi-step operations across multi-domain workflows?

TRON is a reinforcement-learning environment for training enterprise-operations agents on stateful, multi-step work across eight enterprise domains. Each instance embeds a mock enterprise state — HR records, ITSM tickets, calendar entries, drive contents — and a task that can only be solved by choosing the right tool, in the right order, under the right policy. Reward signals are generated from deterministic verifiers over the resulting workflow, decomposing performance into policy adherence, tool-sequence correctness, and end-state fidelity. Difficulty is not asserted — it is measured on frozen signed pilots, with a cohort name and a measured date on every envelope.

  • SQL-based outcome verification: hidden expert-authored SQL queries that validate final database states to ensure exact business outcomes.
  • Resettable containerized sandbox: Model Context Protocol (MCP)-powered live server environments running against freshly seeded SQL database snapshots.
  • Long-horizon stateful planning: multi-step trajectories requiring up to 34 sequential steps with permanent, stateful database mutations.

Three numbers that define the scope of the TRON environment.

Tasks

10,000+

enterprise-ops workflows

Domains

8

HR · ITSM · CSM · Email · Calendar · Drive · Teams · Hybrid

Tools

512

available across domains

How difficulty is measured and rewards are computed. Three principles govern the TRON scoring system.

Principle 01

Dual-Axis Process Verification

  • P-axis (Policy): audits policy adherence — access scoping, approval order, compliance rules, PII handling.
  • T-axis (Tool): audits tool-sequence correctness — the right tool called with the right arguments in the right order.
  • Scores = fraction of passed checkpoints on each axis, per task and per domain.

Principle 02

End-State Fidelity

  • Final environment state is diffed against the golden seed-DB target (HR records, ITSM ticket transitions, calendar mutations).
  • Supported match types: exact record equality, structural subset, field-level tolerance.
  • End-state grading eliminates prompt-parroting: the agent must actually change the world correctly.

Principle 03

Efficiency (Overstep Metric)

  • Overstep = max(0, S_agent − S_human) / (S_human + 1), where S is tool-call count.
  • Penalizes redundant tool calls relative to the human reference trajectory.
  • Reference tool calls per task: — (human baseline).

The method

Three steps from mock enterprise state to scored result.

Phase 01

Author & Seed

  • Task is authored as a JSONL bundle: user prompt, system prompt, expected tool calls, verifiers.
  • Seed DB materializes the mock enterprise state — HR records, ITSM tickets, calendar entries, drive contents — the task will act on.

Phase 02

Act & Comply

  • Agent receives the prompt and a scoped tool set for the target domain(s).
  • Must plan the correct tool sequence and obey the enterprise policy layer (access control, approval order, PII rules).
  • Executes against the containerized, resettable EnterpriseOps-Gym environment via the Harbor adapter.

Phase 03

Verify & Score

  • Process-level verification via dual-axis checkpoints (P-axis policy, T-axis tool).
  • End-state diff compares the resulting environment against the golden seed-DB target.

TODO: results headline (per-model resolution counts).

HR remains the hardest domain across the board — most models sit in the single digits or teens; only — models clear 40%. TRON’s RL environment is designed to close exactly this gap.

Fig. 1: Per-Domain Success Rate (HR, ITSM, CSM, Email, Calendar, Drive, Teams, Hybrid) across evaluated models.

TODO: two-chart section description.

Fig. 2: Dual-Axis Process Scores (P-axis policy, T-axis tool sequence) by domain.
Fig. 3: Failure Mode Distribution (policy violation, tool misuse, sequencing error, end-state drift, hallucinated field, timeout, other).
TODO TODO: label TODO
TODO TODO: label TODO
TODO TODO: label TODO
TODO TODO: confidence strip label
Loading...
Instance Domain Difficulty Task Golden End-State Model A Model B

Head-to-head breakdown of both evaluated frontier models.

Model A

Overall accuracy

Model B

Overall accuracy

TODO: CTA headline.

TODO: CTA sub-copy.

Get in touch