Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
Prime Intellect releases 365,000 unified tasks for RL agent training
The release unifies 23 distinct tasksets under one API.
701.4K127753158.1K
Original post
Developer Impact
Everything installs from a single public GitHub repo and runs on Prime sandboxes
Users can pull the package with one uv command, load any of the 23 tasksets, and execute them inside the company's hosted environments; the release also drops 135,000 prebuilt task images into the public registry.
Open Question
No model runs or accuracy lifts are reported with the new collection
The announcement supplies task counts, validation filters, and access details but leaves open whether any training runs have used the unified set or how performance compares to the original fragmented benchmarks.
Sentiment
Users are excited about Prime Intellect releasing 365,000 tasks for agentic RL environments because the unified API and sandbox remove major research friction and let more people build agents without reinventing infrastructure.
Pos
100.0%
Neg
0.0%
20 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS27.5KBOOKMARKS293LIKES312
a LOT of RL environments (365k tasks) curated in one place, one command 🦋
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
RETWEETS58
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
REPLIES6
We're publishing over 365,000 open and agentic RL Environments for SWE, terminal, and search agents The open research ecosystem has produced many great datasets for the three main agentic domains - software engineering, terminal use, and web research - but every one of them ships with its own harness, its own image conventions, its own grading scripts, and its own failure modes. We integrated them all. 23 tasksets behind one API, one sandbox lifecycle, one command. 365,000+ tasks in total, ~198,000 software engineering tasks across 20+ languages ~28,600 terminal tasks ~137,600 search tasks Ready for evals and RL training on Prime Intellect infrastructure, with validated and cleaned dataset re-uploads where the originals needed fixing.
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
Awesome work! Congrats on the release!
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
~135,000 open-source task images hosted in our own registry, co-located with the sandboxes - to our knowledge the largest such catalog anywhere. A thousand concurrent rollouts, zero Docker Hub rate limits.
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
this is the biggest collection of RL tasks for agentic capabilities
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
Every taskset runs with one command, in any harness, on any runtime - that's the verifiers v1 decomposition doing its job.
~135,000 open-source task images hosted in our own registry, co-located with the sandboxes - to our knowledge the largest such catalog anywhere. A thousand concurrent rollouts, zero Docker Hub rate limits.
the good kind of colocation for agentic RL
~135,000 open-source task images hosted in our own registry, co-located with the sandboxes - to our knowledge the largest such catalog anywhere. A thousand concurrent rollouts, zero Docker Hub rate limits.
at prime intellect, we draw the rest of the fucking owl
the good kind of colocation for agentic RL
The cleaned datasets are re-uploaded with every exclusion and the generation scripts preserved - fully auditable, fully reproducible. All of it lives in our SWE RL collection on Hugging Face.
Read the full post: https://www.primeintellect.ai/blog/scaling-agentic-rl If something you need is missing, tell us or port it: the taskset contract is small, and the sandbox does the heavy lifting.
amazing work by @rasdani_
we've integrated ALL the open source RL envs out there behind one unified API, host the images for you and ran validation+filtering kudos to all the original authors ofc (linked in blog)! we wouldn't be here without your valuable contribs to the open source community 🫶
https://www.primeintellect.ai/blog/scaling-agentic-rl
We're publishing over 365,000 open and agentic RL Environments for SWE, terminal, and search agents The open research ecosystem has produced many great datasets for the three main agentic domains - software engineering, terminal use, and web research - but every one of them ships with its own harness, its own image conventions, its own grading scripts, and its own failure modes. We integrated them all. 23 tasksets behind one API, one sandbox lifecycle, one command. 365,000+ tasks in total, ~198,000 software engineering tasks across 20+ languages ~28,600 terminal tasks ~137,600 search tasks Ready for evals and RL training on Prime Intellect infrastructure, with validated and cleaned dataset re-uploads where the originals needed fixing.
@PrimeIntellect @rasdani_ cooked hard!! 🔥
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
A task is only useful if it produces a clean reward signal: gold patch → tests pass, no patch → tests fail. A surprising amount of open task data fails this. We ran gold-patch and no-op validation on every default dataset - multiple passes, 10x retries to separate flaky from broken.
This drop is significant. We're enabling everyone to be able to access high quality data (+ models + compute) to be able to do agentic RL training at scale. Amazing stuff from @rasdani_
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
This drop is significant. We're enabling everyone to be able to access high quality data (+ models + compute) to be able to do agentic training at scale. Amazing stuff from @rasdani_
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
incredible work by @rasdani_
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
we've integrated ALL the open source RL envs out there behind one unified API, host the images for you and ran validation+filtering kudos to all the original authors ofc (linked in blog)! we wouldn't be here without your valuable contribs to the open source community 🫶
Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.
Load more posts
RL agents share a live sandbox with the grading machinery, and anything readable is fair game for a reward hack. Our integrations withhold test patches, expected outputs, and grading scripts until scoring time - even where the original images ship them readable.