RL Environment Companies for Coding Agents
Patronus AI TeamNovember 202610 min read
Patronus AI publishes this list and is one of the companies on it. Entries are listed alphabetically, not ranked, and each entry links its sources.
The main companies building RL environments for coding agents in 2026 are E2B, HUD, Mechanize, Patronus AI, Prime Intellect, Scale AI, Surge AI and Turing. Some build finished environments from real codebases, some sell ready-made task sets, some give teams a platform or open hub to build their own, and E2B runs the isolated machines underneath. To choose, decide what engineering work your agent must learn, then ask how each task is graded beyond its tests and what the sandbox can run.
A coding agent works the way an engineer does. It reads a codebase, runs commands and submits a change, and it learns in an RL environment that gives it the repository, a task and a reward that scores the result. Grading is the hard part. A patch can pass the tests and still be the wrong change, and a model under training pressure can learn to game the tests instead of fixing the code. This guide examines eight companies serving coding agent buyers in 2026, listed alphabetically. It is written by Patronus AI, a behavioral research lab that publishes at NeurIPS, ICML and ACL.
The eight sit at different layers. Mechanize builds environments and evals for frontier coding agents, and Patronus AI is a research lab that builds environments from real repositories. Scale AI, Surge AI and Turing are data companies that sell coding environments and task sets next to their human data work. HUD is a platform for building environments in-house, and Prime Intellect runs an open hub with a coding section plus sandboxes and hosted training. E2B supplies the isolated machine each agent session runs on.
Last updated
Key takeaways
- Start from real repositories, not puzzles. Tasks that teach engineering live in codebases with history, tests and undocumented decisions. Turing frames its catalog as engineering work rather than coding puzzles, and Scale AI’s SWE Atlas runs agents inside environments built from real software repos.
- Treat passing tests as the floor. A change can pass every test and still be wrong, and agents can learn to edit tests or hardcode outputs. HUD flags test manipulation, output hardcoding and scorer exploits, and Scale AI pairs tasks with process as well as outcome verifiers.
- Look past bug fixes. Engineering also means building features, deploying, reviewing code and tuning performance. Mechanize’s tasks include deploying an application, and Turing covers systems engineering, security, QA and code review.
- Match the sandbox to the workload. Short code execution runs fine in containers, but tasks that need Docker inside the sandbox, background services or kernel features need a full machine. Prime Intellect’s sandboxes are full Linux VMs with native Docker, and E2B runs a microVM per agent session.
- Ask how the grader itself was checked. A grader that passes broken work teaches the wrong lesson at scale. HUD runs QA agents that find tasks whose verifiers cannot be trusted, and Turing grades its GPU kernel tasks against a speedup a human solution already reached.
Comparison table
All eight companies at a glance, listed alphabetically. Security shows what each company lists on its own site. Sort by any column marked with arrows.
| Company | Best for | Type | What it sells | How tasks are graded | Security |
|---|---|---|---|---|---|
| E2B | Sandboxes for teams that bring their own tasks | Infrastructure | Isolated microVM sandboxes for agents and RL rollouts | Set by each environment | SOC 2 Type II, HIPAA |
| HUD | Building coding environments in-house | Platform | SDK, coding template and Harbor task import | Hidden tests, with agents auditing every trace | SOC 2 Type I (Type II in progress) |
| Mechanize | Software engineering work from a coding specialist | Specialist | Environments and evals for frontier coding agents | A grader scores each run | Not listed on its site |
| Patronus AITalk to us → | Undocumented codebases, from a research lab | Research lab | Environments built from real repositories, ready-made and custom | Tests plus reviewer checks, every task reviewed by experts | SOC 2, HIPAA, TISAX |
| Prime Intellect | Open environments and a full training stack | Open hub | Environments Hub with a coding section, sandboxes and hosted training | Open-source verifiers library | Not listed on its site |
| Scale AI | Coding next to tool use and computer use, from one large vendor | Data and evaluation | RL environments across tool use, computer use and coding, plus SWE Atlas | Process and outcome verifiers on expert-designed tasks | SOC 2 Type II, ISO 27001, FedRAMP High |
| Surge AI | Ready-made repository tasks | Expert data | Off-the-shelf repository tasks and custom RL environments | Expert-designed verifiers | Not listed on its site |
| Turing | The whole software development lifecycle | Expert data | Software engineering environments and datasets, off the shelf and custom | Engineer-authored tasks and rubrics | SOC 2 Type II |
E2B
E2B runs the isolated machine an agent works on, in its cloud or in a team’s own AWS, GCP or Azure account. It lists coding, computer use, background jobs and RL rollouts as uses of one API.
How does E2B work?
- One isolated microVM per agent session, on a runtime built on Firecracker with dedicated kernels.
- For reinforcement learning, thousands of machines spin up, grade and clear.
- The runtime is open source under Apache-2.0, and E2B says it has run 2B+ sandboxes.
Where it fits
E2B supplies the machine, not the tasks or the grader, so teams pair it with environments they build or buy.
Best for: Teams that have coding tasks and graders and need isolated machines to run rollouts at scale.
Sources: e2b.dev
HUD
HUD is a platform for building RL environments. Teams start from a community environment, build their own with its SDK, or import existing tasks from Harbor.
How does HUD work?
- Its coding example is a complete repository environment. The agent works in a sandboxed workspace while the environment keeps a clean baseline outside it, then applies the authored tests after the agent finishes.
- Teams with coding tasks in Harbor format can import them and let QA agents find the tasks whose verifiers cannot be trusted.
- Agents audit every trace for grader mistakes and reward hacking, and a detector flags test manipulation, output hardcoding and scorer exploits.
Where it fits
HUD suits teams that want to write coding tasks with their own engineers on a shared platform, with trace audits built in.
Best for: Teams building their own coding environments who want reward hacking checks on every run.
Sources: docs.hud.ai/v6/cookbooks/coding-agent.md, trust.hud.ai
Mechanize
Mechanize builds environments and evals for frontier coding agents.
How does Mechanize work?
- Models carry out software engineering work such as building a feature, deploying an application or debugging an issue in an unfamiliar codebase.
- A grader scores each run, and the scores serve as signals during reinforcement learning and evaluations.
- GBA Eval, its public benchmark, measures how well coding agents can write a Game Boy Advance emulator from scratch in 24 hours.
Where it fits
Mechanize argued in an August 2025 essay that labs would come to favor quality over quantity when buying RL environments. Its site does not say whether it sells ready-made environments or custom builds.
Best for: Teams that want software engineering tasks from a company focused on coding agents.
Sources: mechanize.work
Patronus AI
Patronus AI is a behavioral research lab for AI. It builds RL environments for coding agents that start from real repositories, alongside environments in alignment, knowledge work, science, computer use and more.
How does Patronus AI work?
- Real repositories. Each environment starts from a real codebase, with its file tree, commit and pull request history, tests, docs and run commands. The task does not say where to look.
- Assigned work, from small bug fixes to features that touch the database, the API and the frontend. Experts review every task before it is used.
- Grading beyond tests. Tests are the starting point. Where a task needs more, Patronus AI adds the checks a reviewer would make, written with engineers who know that kind of codebase, and looks at the process as well as the result.
- Difficulty set against the models in training, and retuned as new models arrive.
Where it fits
Patronus AI publishes on how models and agents behave and where they fail. Its ICML 2026 paper Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis sorts reward exploits in code RL into 54 categories across 517 trajectories, and finds models catch more of them side by side than one at a time, 63% for the best model tested against 45% in isolation. In MEMTRACK: Evaluating Long-Term Memory and State Tracking in Multi-Platform Dynamic Agent Environments, the facts an engineer needed were spread across Slack, Linear and Git, and the best model tested answered 60% correctly. Ready-made environments and custom builds are both offered, and most work is custom.
Best for: Teams training coding agents on real, underspecified engineering tasks who want graders checked against reward hacking.
Sources: patronus.ai/domains/coding, patronus.ai/research/trace
Prime Intellect
Prime Intellect runs an integrated compute, training and inference stack, and lists coding agents among its uses.
How does Prime Intellect work?
- Teams build RL environments with its open-source library or start from 2,500+ community environments on its Hub, which has a coding section with long-horizon software engineering tasks.
- Rewards are written with the open-source verifiers library.
- Its sandboxes are full Linux virtual machines with native Docker, background jobs and kernel-dependent workloads (Sep 23, 2026).
Where it fits
Prime Intellect suits teams that want open environments and their own training runs on one stack. Its homepage offers serverless or VPC deployment, role-based access control and audit logs; its site lists no certifications.
Best for: Teams that want open coding environments, sandboxes and hosted training in one place.
Sources: primeintellect.ai, app.primeintellect.ai/dashboard/environments
Scale AI
Scale AI provides data, evaluations and RL environments to AI labs, governments and enterprises. Its Feb 27, 2026 launch post describes environments for tool use, computer use and coding.
How does Scale AI work?
- Tasks designed by domain experts, each paired with process and outcome verifiers.
- SWE Atlas, an evaluation suite, runs agents inside reproducible environments built from real software repos, with tasks written by professional engineers.
- Every RL environment it builds runs on AgentEnv, an open-source framework released Oct 5, 2026.
Where it fits
Its RL environments product page now centers on enterprise systems and desktop environments, so ask which coding environments are available. Its security page lists a SOC 2 Type II report, ISO/IEC 27001:2022 and FedRAMP High authorization.
Best for: Teams that already buy data or evaluations from Scale AI, or that need FedRAMP High.
Surge AI
Surge AI sells training data, evals and RL environments built by the experts behind its lab work, off the shelf or as custom programs.
How does Surge AI work?
- Off-the-shelf coding tasks set inside complete codebases, covering feature development, bug fixes and production refactors, compatible with SWE-bench out of the box.
- Tasks built from real company codebases that reflect the full software development lifecycle.
- Custom programs include reinforcement learning and custom RL environments, with verifiers designed to reward agent behavior.
Where it fits
Surge AI suits teams that want a ready-made set of repository tasks or a custom program with a large data vendor.
Best for: Teams that want off-the-shelf coding tasks that work with SWE-bench tooling.
Sources: surgehq.ai/ots, surgehq.ai/enterprise
Turing
Turing builds RL environments and datasets for AI labs, including software engineering environments designed to capture the end-to-end development lifecycle.
How does Turing work?
- Long-horizon tasks across coding, systems engineering, security, QA, code review and more, framed as engineering work rather than coding puzzles.
- A catalog of off-the-shelf and custom datasets, including CyberStrike with 200 tasks and KernelQuest, where each GPU kernel task is graded against a speedup a human solution already reached.
- Harnesses capture trajectory traces automatically (RL environments).
Where it fits
Turing suits teams that want coverage beyond bug fixes, such as security, review and performance work. Its data program site lists SOC 2 Type II, with the report available under NDA.
Best for: Teams that want software engineering tasks across the whole development lifecycle.
Sources: turing.com/frontier-ai/software-engineering, lazarus.turing.com
Where Patronus fits
Patronus AI wrote this guide and is one of its entries. It builds coding environments from real repositories, with tasks reviewed by experts and graders that add a reviewer’s checks to the tests, alongside environments in other domains. Ready-made environments and custom builds are both offered, and it confirms how tasks run on a team’s own training harness when it scopes the work.
Its research on coding agents includes Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis at ICML 2026, which released its benchmark and evaluation harness. Its researchers also train models on their own tasks and check passing runs by hand (how it was measured). It lists SOC 2, HIPAA and TISAX, with data ownership set in each agreement. See how it builds coding environments.
Frequently asked questions
What is an RL environment for coding agents?
It is a sandbox with a codebase, a task and a reward. The agent explores the repository, runs commands and submits a change, and the grader scores the result, usually starting from tests. Our page on coding RL environments covers how Patronus AI builds them.
Which companies build RL environments for coding agents?
E2B, HUD, Mechanize, Patronus AI, Prime Intellect, Scale AI, Surge AI and Turing, among others, listed here alphabetically. The table above sets out what each one sells and who it suits.
How are coding tasks graded?
Most start from tests. HUD applies hidden tests after the agent finishes, Scale AI pairs tasks with process and outcome verifiers, and Patronus AI adds the checks a reviewer would make where tests are not enough.
What is reward hacking in coding environments?
It is when a model earns the reward without solving the task, for example by editing the tests or hardcoding the expected output. HUD flags these patterns in traces, and Patronus AI’s ICML 2026 paper Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis catalogs 54 categories of them.
Do I need a separate sandbox provider?
Only if you build your own environments. E2B and Prime Intellect run isolated machines for rollouts, and Scale AI’s open-source AgentEnv supports E2B and Modal sandboxes. Companies that sell finished environments either host them or deliver them to run on your own setup, so ask which.
Where can I compare RL environment companies across domains?
Our guide to RL environment companies covers the wider field, including expert data companies and open hubs.
How was this list made?
Patronus AI wrote it and is one of the companies on it. The entries are listed alphabetically, not ranked, and each entry links its sources so you can judge for yourself.
Last updated