Agentic Tasks & RL Environments

If you can take a messy real task, pin down exactly what "done correctly" means, and turn that into a check a machine can run thousands of times, you are qualified for the best-paid and fastest-growing corner of the entire AI-training economy. Building the interactive tasks and reinforcement-learning (RL) environments that agentic models train against sits at the top of the pay ladder for one blunt reason: it is not labeling, it is software and task engineering, and the frontier labs are spending on it at a scale that dwarfs the rest of the field. Anthropic alone was reported to be discussing $1 billion or more in a single year on environments, and an industry tally circulated in July 2026 counts 50-plus vendors selling data and RL environments to labs at roughly $8.5 billion in combined annual revenue. This guide covers what the work actually is, who buys it, what it really pays per task and per hour, what you genuinely need to do it, and the open on-ramp that lets anyone in the world build a portfolio the labs can see.

The honest headline up front is that this is a two-sided market. At the top, the startup Mechanize reportedly offered $500,000 salaries to engineers who build RL environments, and single authored tasks change hands for $200 to $2,000 each, with a full product replica reaching six figures. At the bottom, the national-average AI-trainer freelance rate is about $31 an hour and entry labeling still pays $12 to $25. The gap between those two worlds is precisely what this guide is about: the skill of designing tasks and writing verifiers, not the willingness to click.

What "Agentic Tasks" and "RL Environments" Actually Are

Modern models are increasingly trained to act, not just answer. They run code, click through applications, search files, and complete multi-step tasks. Teaching that behavior needs a fundamentally different human input than rating two chatbot answers. According to Epoch AI's authoritative FAQ on the subject, an RL environment has three parts: a set of actions the model can take (run code, call a tool, edit a file), the surrounding state that determines what those actions do (a file system, an application, a simulated app), and a grader that scores each attempt. Scored attempts update the model's weights so it repeats what worked. One founder described building one to TechCrunch as "creating a very boring video game": a small, rule-bound world with a win condition.

RLVR: the shift from preference to proof

The paradigm underneath this is RLVR, reinforcement learning with verifiable rewards. Where classic RLHF trains a learned reward model from human preferences (a person picks the better of two answers, and a second model learns to imitate that taste), RLVR gives the model a reward only when its output passes an automated, deterministic check: the unit tests pass, the math answer matches, the database ends in the correct state. As Appen explains, the reward is computed, not predicted. The advantage is structural: there is no learned reward model to hack, because a deterministic verifier cannot be flattered. RLVR started where answers are cleanly checkable (math, code, logic) and in 2026 is being pushed into agent workflows (tool use, computer use, enterprise tasks) where, as Daily Dose of DS notes, writing a trustworthy verifier is much harder and the human judgment is worth much more.

The vocabulary you need

A handful of terms recur across every listing, and knowing them is the difference between reading a job post and understanding it:

• Trajectory (or rollout). One full attempt by the agent: its complete sequence of actions and observations. Labs generally want tasks with a 2 to 3 percent pass rate, roughly one success per 64 to 128 rollouts. A task the model never solves may be broken or infeasible; a task it always solves teaches nothing.
• Grader (or verifier, or rubric). The code or expert-written rule that scores the attempt. Robustness against reward hacking was cited as the single most important quality criterion by the environment founders Epoch surveyed, because a model will find and exploit any loophole in a sloppy check.
• Gym. A colloquial name for an interactive environment. Epoch reports that OpenAI bought "hundreds of replicated user-interface gyms" to train agents to operate software.
• verifiers / prime-rl. Prime Intellect's open-source library and trainer for building environments with built-in reward functions, and the most accessible way for an individual to actually author one (more on that below).

Where the demand is concentrated

Epoch groups the demand into three domains: math and coding (the original focus, with math emphasis now declining), enterprise workflows (the major growth area: expense reports, spreadsheets, CRM navigation, multi-tab office tasks), and longer-horizon multi-step tasks. Coding is widely treated as the "holy grail" vertical, and the messy enterprise-workflow environments are the fastest-growing, which matters for where a newcomer should aim.

What You Actually Do, Day to Day

Not all of this work requires you to engineer a sandbox from scratch. The category spans a ladder from an accessible on-ramp to deep specialization, and most contracts mix several of these tasks. Per Epoch and the Wing Venture Capital market analysis, contributors design tasks, write the reference solution, specify the reward signal, containerize the environment, and in some cases grade the model's rollouts and QA the whole thing against reward hacking.

The through-line is the same across the ladder: your product is a decision about what correct means and a mechanism that enforces it. A trajectory annotator decides whether a given step was a mistake; an environment builder decides what the whole task even rewards. Both are judgment about correctness, encoded so a machine can act on it repeatedly. That is why the work resists automation and why it pays what it does.

To make it concrete, consider a coding environment built end to end. You take a real bug in an open-source repository, write the task ("make the failing test pass without breaking the others"), and set the grader to the existing test suite itself, so pass or fail is decided by code rather than opinion. You containerize the repo so an agent can check out the code, edit files, and run the tests in a sandbox, then you tune the difficulty until the model succeeds on roughly one attempt in fifty, the band Epoch says labs want. Finally you attack your own grader: can the model make the test pass by deleting it, hard-coding the expected output, or weakening an assertion? Closing those holes is the actual craft, and it is why a well-built environment is worth far more than a raw bug report.

Who Buys It, and What Shape the Market Is In

The buyers are the frontier labs (Anthropic, OpenAI, Google, Meta, xAI) and a growing set of product companies. Some, such as Cursor, build environments in-house; others, per Epoch, partner with labs on replicas of their own products (Salesforce, Slack, Shopify, and Stripe among them), so a model can learn to operate the real software.

The vendor market has ballooned. The July 2026 tally counts more than 50 companies selling data and RL environments to labs, roughly $8.5 billion in combined annual revenue and about $100 billion in combined valuation, with more than 75 percent concentrated in four players: Scale, Surge, Mercor, and Handshake. For context, Wing VC puts the older, traditional labeling market at about $5 billion growing more than 50 percent a year, and predicts the roughly 20 seed-to-Series-A RL-environment startups will consolidate to 3 to 5 leaders by 2030. Treat the $8.5 billion and $100 billion figures as an analyst estimate rather than an audited number, but the direction is not in dispute.

The vendors, by verified status (mid-2026)

• Mercor pivoted hard from talent marketplace into RL environments and domain-expert data. It raised a $350 million Series C at a $10 billion valuation in October 2025, and Sacra estimates it reached roughly $2 billion in annualized gross revenue by June 2026. In February 2026 it acquired Sepal AI, a platform building training data, eval benchmarks, and RL environments for frontier LLMs, to "train models against the most complex tasks humans perform."
• Surge AI is bootstrapped and large: about $1.2 billion in 2024 revenue, with reporting of a first external round of up to $1 billion at a roughly $24 billion valuation (reported, not closed). It built a dedicated internal RL-environments team and ships EnterpriseBench, a simulated startup world where frontier models solve only about 30 percent of enterprise agentic tasks and fail more than 40 percent of simulated workplace tasks, a number that quantifies the entire opportunity.
• Scale AI / Outlier is the cautionary tale. After Meta took a roughly $14.3 billion stake in June 2025, rival labs cut contracts and Scale laid off about 14 percent of staff. Its CEO says 2026 revenue is expected to exceed just $1 billion, down from about $2 billion in 2025. That decline is the clearest single illustration of the shift away from static labeling.
• Prime Intellect is the open on-ramp. It launched its Environments Hub in August 2025, a community platform that now hosts more than 2,500 crowdsourced RL environments, and it ships the open-source prime-rl trainer and verifiers library. It runs an Environments Bounty Program paying $1,000 to $5,000-plus per environment, open worldwide.
• Turing repositioned around "AGI advancement," raising a $111 million Series E at a $2.2 billion valuation, and built a 10,000-plus Python-task dataset for reinforcement learning with execution feedback.
• Handshake AI pivoted from a college-jobs platform to AI data and reached about $1 billion in gross annualized AI-training revenue by April 2026, one of the big four, acquiring the quality-tooling firm Cleanlab to move into evaluation and QA.
• Micro1 went from roughly $7 million to about $200 million annualized in 2026, growing around 35 percent a month, on the strength of its AI recruiter, and is reported at a $2.5 billion valuation.
• Mechanize, founded in April 2025 by ex-Epoch researchers to build RL environments for full economic automation, is working with Anthropic and offered $500,000 salaries to environment engineers; Tracxn reports about $9.1 million raised at a roughly $500 million valuation (reported).
• Invisible Technologies raised about $100 million in September 2025, earmarking funds for RL gyms and "mirror worlds," digital twins of real operations.
• Alignerr (over Labelbox) is the contributor-facing marketplace, with community-reported pay around $15 to $60 an hour and up to $25 to $125 for credentialed specialists.

Newer pure-play entrants that Wing names include Bespoke, Applied Compute, and Cognition, and directories such as rl-list.com track the growing field.

What It Pays

Pay in this category splits into two very different numbers: what the lab pays the vendor per task, and what the vendor pays you per hour. Understanding both keeps you from either underselling your work or believing a headline that is really an enterprise contract price.

Per-task pricing (what labs pay vendors)

Per Epoch AI, individual expert-authored tasks run $200 to $2,000 each (a rare, very hard one can hit $20,000). A website replica costs on the order of $20,000, and a complex product replica such as a Slack clone can reach $300,000. Exclusive deals cost roughly 4 to 5 times a non-exclusive one, and full lab contracts are often six to seven figures per quarter. These are vendor-level prices, not your paycheck, but they explain why the hourly rates below are high: the underlying work is valuable.

Hourly pay to the worker

Sources: Mercor's own rate guide puts the national average AI-trainer rate at $31/hr, with expert and credentialed work at $75 to $200+, and CEO Brendan Foody stated the platform's average pay rate "surpassed $100/hr". Software engineers building RL environments on Mercor sit roughly in the $70 to $160/hr band (the exact figure on a specific listing is best confirmed live). Domain experts (PhDs, physicians, lawyers, senior engineers) commonly earn $75 to $200+/hr, with the top medical, legal, and finance tiers reported at $150 to $300+/hr, and PhDs and grad students are described as the highest-paid tier by a wide margin. Separately, Prime Intellect's bounties pay $1,000 to $5,000-plus per environment, which behaves more like piecework and portfolio-building than an hourly job.

The honest caveat: the $500,000 salaries and $300/hr tiers are the top of a wide distribution, not the median. Plan around the expert band and treat the headline as the ceiling.

What a realistic engagement looks like: the expert marketplaces quote environment-engineering work at 15 to 25 hours a week, flexible up to 40, on contracts that run at least a quarter. A credentialed engineer holding steady expert-tier work at, say, $110 an hour for 20 hours a week grosses roughly $2,200 in a full week, though like all contract work the volume is lumpy and the effective rate sits below the headline once you count ramp-up and unpaid calibration. Bounty work behaves differently again: a single strong environment can pay $1,000 to $5,000-plus as a one-off, which makes it a portfolio investment as much as an income stream.

What You Need to Do This Work

This is where agentic-eval work separates cleanly from generic labeling. The premium roles require real software engineering: writing Python, building containerized environments with Docker, and authoring unit-test-style graders that resist reward hacking. Because reward functions are custom code, you often cannot tell whether the function measures what you think until after training, so debugging and verification skill matter as much as the initial build.

Alongside the engineers, labs pull in domain experts to define domain-specific tasks, write reference solutions, and specify reward signals: finance professionals for financial tasks, doctors for medicine, lawyers for legal reasoning. Machine-learning literacy helps but is not always required for domain-expert authoring, because the labs supply the training machinery; what they cannot supply is your judgment about what a correct outcome is. The contrast with RLHF preference-rating is sharp: rating needs consistent judgment against a rubric, while environment authoring needs the ability to build and test executable software with a correct automated grader, or the equivalent expert reasoning trace in a licensed domain.

In practice the single most valuable engineering skill here is verifier debugging. Because a reward function is just code, a subtle mistake can quietly reward the wrong behavior for an entire training run before anyone notices, which is exactly the failure mode that made robustness against reward hacking the top-cited quality criterion. Engineers who can write a grader, adversarially test it against a model that is actively trying to cheat, and then confirm after training that it measured what they intended are precisely what the labs are short of.

How to Break In (Including the Open On-Ramp)

Unlike most premium work, this category has a genuinely open door that requires no application at all.

1. Build a public environment on Prime Intellect. The Environments Hub and Bounty Program let you build and publish an environment using the open-source verifiers library, earn $1,000 to $5,000-plus, and create a public artifact a lab can inspect. Prime Intellect crowdsourced more than 400 frontier environments in about two months this way. HUD.ai is another route to publish environments to labs.
2. Apply to the expert marketplaces. Mercor (an AI-vetted interview), Surge AI, Handshake AI, Turing, Micro1, and Alignerr all recruit for this work; browse the live openings above for what is posted right now.
3. Expect a real screen. A technical or coding assessment for engineering roles, credential verification for domain experts, then trial tasks. Contracts typically run at least a quarter, part-time or full-time (Mercor's SWE listings quote 15 to 25 hours a week, flexible to 40).
4. Lead with the right portfolio signal. A published RL environment on GitHub or the Environments Hub, demonstrated Python plus containerization plus a clean grader that resists reward hacking, and domain credentials for the specialist tracks.

The strategic point: the on-ramp and the credential are two different doors. A strong engineer without a PhD gets in through built artifacts; a physician or lawyer gets in through the credential and a reasoning trace. Aim at the door that fits what you already have.

Common Misconceptions

• "It is just the next word for labeling." No. Commodity labeling is being automated and its rates are falling, while designing tasks, graders, and reasoning traces is priced at $200 to $2,000 per task and $75 to $300+/hr. It is the opposite end of the same market.
• "You need a PhD." Only for the credentialed tracks. Domain specialists command $150 to $300+/hr, but strong software engineers without a PhD are in high demand to build environments, and the Prime Intellect bounties are open to anyone worldwide.
• "The work is drying up." Not this category. Static labeling is contracting ( Scale's 2026 revenue roughly halving) while RL environments are the growth area, with Anthropic discussing $1 billion-plus a year.
• "AI will just build its own environments." That is Mechanize's long-term thesis, but 2026 demand for human-built environments is rising, not falling, because the bottleneck is quality at scale, not availability of tooling.

The 2026 Outlook

Frontier labs are spending heavily because RL environments are the binding constraint on the next capability jump. Models are being trained to act, and that requires interactive, verifiable environments rather than static text. Epoch reports that RL is now the dominant use of this data (10 to 20 times more demand than benchmarking or supervised fine-tuning) and that enterprise-workflow environments are the fastest-growing domain. The stated "number one bottleneck" is maintaining quality while scaling: contractor motivation, verification, and anti-reward-hacking QA, all of which keep human labor central even as tooling improves. Prime Intellect frames closed, expensive environments as the reason open models fall behind, which is the motivation behind its open Hub.

The realistic forecast for a worker: expect continued consolidation ( Wing sees 3 to 5 winners by 2030), continued lab in-sourcing, and rising demand for people who can build and verify environments, especially in coding and messy enterprise workflows. The floor of generic labeling will keep eroding; the ceiling for people who can specify correctness will keep rising.

Bottom Line

Agentic tasks and RL environments are the highest-paid, fastest-growing, and least understood corner of AI-training work in 2026. If you can engineer a sandbox and write a grader that a model cannot cheat, or if you hold real domain expertise and can author the reasoning a model should learn from, this is the most durable and best-compensated work in the field, running from $75 to $200+/hr for experts up to authored tasks worth $200 to $2,000 each. The open door is real: build one environment on Prime Intellect's Bounty Program, publish it, and you have an artifact the labs can see. If your background is generalist today, treat this as the direction to grow toward: start with rating and trajectory annotation, build genuine depth in one domain, and move up into graders and environments as your track record grows.