● RL Studio · private beta
Real workflows become RL environments: simulated systems, typed tools, and verifiers your agents can’t game.
$ harbor rl reset env=legal task=liability_cap_conflict seed=4471
Fix a liability cap that breaches client policy
-.--
validation gates before an environment ships
simulated system types, from documents to databases
lines of your code we need to run
zip to run it on your own infrastructure
[01] Reward you can trust
Models have read the internet. They haven’t done your job. An environment lets them practice it thousands of times, and every grader is attacked with known exploits before it’s allowed to publish.
Verifiers
Reward hacking
Determinism
[02] How it works
Describe the workflow
/01
Describe the workflow
Write it the way you’d brief a new hire. RL Studio drafts the systems, actors, tools, tasks, and verifiers, and asks about anything it can’t infer.
/02
Map your tools
Import an OpenAPI spec or an MCP server, or define tools by hand. Every endpoint becomes a typed action against a simulated system, never your production one.
exploit · “approve own request” → reward 0 · blocked
/03
Validate
Eight gates check the spec, the runtime, determinism, and the verifiers. Every task ships with exploit traces, and an environment that pays out for one doesn’t publish.
Success
0%
Avg reward
0.20
Critical
0
/04
Evaluate
Connect your agent by webhook, or run a Claude baseline. Every episode is traced step by step and can be replayed exactly.
$unzip contract-review-v3.zip && cd contract-review
$docker build -t contract-review . && docker run -p 8000:8000 contract-review
listening on :8000 · engine 1.0.0 · checksums ok
>>>env = HarborEnv("http://localhost:8000")
>>>env.reset(seed=7)
>>>env.step("read_document", {"id": "msa_v3"})
StepResult(reward=0.0, done=False, ...)
/05
Export
Download one zip with the engine, your environment, a Dockerfile, and a Python client. Run it in your own training stack, with or without Harbor.
[03] Published results
Every catalog task runs on every variant: the correct solution, wrong answers, every exploit we wrote, and attack agents that know nothing about the task. The numbers regenerate whenever an environment changes.
7,903
episodes behind these numbers
397
of 397 exploit strategies blocked
128
held-out evaluation variants
456
of 456 replays matched hash for hash
| Policy | Expected | Episodes | Success |
|---|---|---|---|
| Reference solution | should pass | 456 | 100%✓ |
| Wrong answers | should fail | 528 | 0%✓ |
| Authored exploits | should fail | 1,191 | 0%✓ |
| Memorized answer | should fail | 256 | 0%✓ |
| Submit immediately | should fail | 456 | 0%✓ |
| Call every tool | should fail | 456 | 0%✓ |
| Random agent | should fail | 4,560 | 0%✓ |
[04] Environment catalog
25 validated environments, 150 tasks and 450 seeded instances across 9 sectors. Not listed? Any workflow that runs through software and has a definition of done can become one, and agents never touch your production systems.
Legal
Verifiers
[05]The studio
[06] Two ways to work with us
Service 01
Your team builds environments in RL Studio: describe the workflow, import your APIs, validate, evaluate your agent, and export. Start from validated templates across legal, banking, insurance, healthcare, bio, industrial and more.
Service 02
Harbor’s team and expert network build them with you: practicing lawyers, analysts, and operators write the tasks, edge cases, and rubrics, grounded in real workflow data licensed through Harbor Enterprise.
[07] Buyer’s questions
You’re buying a reward signal as much as a simulator.
A simulated version of a workflow that an agent can practice in: the systems it touches, the tools it can call, the tasks it’s given, and verifiers that score how well it did. Agents improve by running many episodes and learning from the rewards.
No. Environments are specifications executed by Harbor’s own engine. You describe systems, tools, and checks; nothing you upload is executed, so there is no sandbox to escape.
No. Imported APIs become simulated systems with the same shapes. Agents act on the simulation, and your live systems are never called.
Every task ships with exploit and negative traces. The verification gate runs them before an environment can publish and fails it if any shortcut earns reward. Critical verifiers zero the reward, and agents never see the grader. We publish the full attack record with our results.
No. Rewards come from deterministic checks on the simulated systems. Expert rubrics are routed to human review and never auto-scored by a model.
Yes. Each catalog task keeps held-out variants for evaluation. They are never served in the training split, never exported and never published, and a memorized training answer solves none of them.
Any agent that can receive a signed webhook, plus an OpenEnv-style API (reset, step, state) for training loops. Exports include a Python client and a zero-dependency Node server in a Docker image.
Yes. Every validated environment exports as one zip with the engine, your spec, a Dockerfile, a Python client, and checksums. The same version always produces the same archive.
From your workflow description and APIs, and, if you want, from real workflow records and expert-written tasks through Harbor Enterprise and our expert network.
With an API key. List environments with GET /catalog, then reset and step any task over the hosted OpenEnv-style API, or with our Python client, from your own training loop. Run episodes in parallel, pull full trajectories for audit, or export an environment and run it on your own infrastructure.
You do. Everything your agents generate in an environment belongs to you, and Harbor doesn’t train on it or share it. Exclusive licences are available.
Catalog environments are licensed per environment, and custom environments are scoped and priced per build. Both are invoiced under a contract, with API keys, the hosted runtime and exports included. Most teams start with a paid pilot on a few environments. Request access and we’ll scope one with you.
[08] Request access
Tell us what you’re training. We’ll set up your workspace and build a first environment from one of your workflows with you. Agents can also buy access directly through the Harbor MCP server.
See it on your workflow
We build a first environment from one of your workflows with you.
Baseline your agent
Find out where your agent fails before you train it.
Check the grader
Watch exploit traces get caught before anything ships.
Take it with you
Leave with an export you can run in your own stack.
