pretty_name: MartechBench
language:
- en
tags:
- rl-environment
- tool-use
- marketing-automation
- commercial
- harbor
- verifiers
- enterprise
- agents
- long-horizon
viewer: false
MartechBench
Enterprise work, modeled as deterministic rewards for LLM post-training.
A commercially licensed RL environment for frontier labs and research teams training agents to operate real business workflows. In the enterprise automation research category alongside AutomationBench, MartechBench focuses on the connected marketing stack: durable state, dependencies, realistic workplace context, and measurable downstream consequences.
1,400+ tasks · 400+ tools · Deterministic verifiers · Multiple valid solutions per task
Discuss licensing with ReadySolve →
A change in one system changes the next
Agents must resolve ambiguous requests, find the current owner, and change upstream state in dependency order. Success depends on what happens downstream—and on preserving the workflows around it.
The runtime models connected products, persistent records, permissions, versioned configuration, and time. Tasks demand investigation, coordination, and execution across system boundaries. Verifiers inspect the business state the agent leaves behind; a plausible final answer is not enough.
Built for outcomes, with room for different solutions
Each task supports multiple valid solutions. Deterministic verifiers grade durable business outcomes rather than a prescribed tool sequence or a single golden configuration. Partial credit captures meaningful progress; protected neighboring behavior and held-out probes test whether the workflow actually works.
Quality built into every task
Every task is versioned and backed by deterministic outcome checks. Rigorous review and repair strengthen the task instructions, workplace evidence, and verifiers. Researchers receive an exact, reproducible taskset ready for evaluation and post-training.
Nearly 80% of tasks have three or more versions. This depth of review and refinement delivers a carefully hardened corpus, with versioned tasks and deterministic verifiers that researchers can pin and reproduce.
Ready for your research infrastructure
| Research use | Delivery |
|---|---|
| Capability evaluation and model comparison | One canonical Harbor-native buyer release |
| RL and post-training experiments | Execution in buyer-controlled infrastructure |
| Reproducible experiments | Pinned task revisions, runtime, verifiers, and artifact hashes |
| Internal auditing and retention | Licensed build materials and container artifacts |
Corpus scope is approximately 1,400 tasks; the current runtime exposes 426 tools. Exact purchased membership and framework compatibility are specified in each licensed release. Additional integrations consume the same release rather than creating separate editions.
Commercial access
MartechBench is proprietary research infrastructure, available through a commercial agreement. Reach out through readysolve.sh →
This public page contains product information and illustrative visuals only. Task data, runtime code, and verifiers are available to authorized buyers. Task production, internal QA machinery, private rollout evidence, and release curation remain internal.
