Prime Intellect

Agents built in Prime Intellect environments reach outside systems through the tools their harness gives them: environments declare tools, sandboxes and stateful tool environments, hosted evaluations can be granted temporary sandbox, instance or tunnel permissions for tool-using environments, and Prime Tunnel exposes local services to them. No native connectors, OAuth actions or credential scoping for business systems are documented, so tools are code the customer writes into the environment.

Sourcedocs.primeintellect.ai/hosted-training/what-is-labread 2026-09-22

Lab runs agent steps in a loop the customer defines: an environment's harness sets the model's tools, sandboxes, context management and multi-turn interaction, and for each Hosted Training run a dedicated Orchestrator manages environment logic, schedules rollout requests and coordinates training; multi-node clusters can be orchestrated with Slurm. The harness and rollout orchestration serve training and evaluation and are written in code. No workflow runtime is documented that sequences, branches, retries and routes a deployed agent's production work.

Sourcedocs.primeintellect.ai/hosted-training/what-is-labread 2026-09-22

An environment's dataset is a set of tasks with optional answers for training and scoring, not the customer's knowledge for an agent to retrieve, and the search-agent guide has the customer build its own document search inside an environment. No retrieval structure that Prime Intellect maintains over the customer's content is documented.

Sourcedocs.primeintellect.ai/hosted-training/what-is-labread 2026-09-22

No approval step, consent checkpoint or escalation to a person is documented for the agents that run in Prime Intellect's environments or sandboxes. The closest control is sandbox egress: network allow and deny lists restrict where an agent's code can reach, but that is isolation the developer configures per sandbox, not a point where a person reviews the agent's action.

Sourcedocs.primeintellect.ai/sandboxes/overviewread 2026-09-22

Sandboxes are hardware-isolated microVMs with network allow and deny lists for egress, secrets are injected encrypted and never appear in logs or API responses, API keys carry scoped permissions (Inference, for example), and team members hold Admin or Member roles. Identity is the gap: no SSO, SCIM or audit log is documented, and no SOC 2 or ISO attestation is published.

Sourcedocs.primeintellect.ai/sandboxes/overviewread 2026-09-22

Prime Intellect keeps a record of runs: hosted evaluations store each run with its results in Prime Evals, stream logs to the CLI and dashboard, and can be shared; Hosted Training reports run metrics; reserved clusters come with Grafana dashboards and alerting on GPU, node and network health. Run logs and results are there, but no audit log, step by step trace retention or SIEM export is documented.

Sourcedocs.primeintellect.ai/tutorials-environments/hosted-evaluationsread 2026-09-22

No memory an agent reads across runs is documented. What persists is compute state. Sandbox snapshots that save, restore and fork a sandbox mid run are marked coming soon and have not shipped, paused instances keep a workspace directory, and persistent disks hold datasets and checkpoints. All of it is storage the platform applies, not context the agent reads to decide.

Sourcedocs.primeintellect.ai/sandboxes/overviewread 2026-09-22

Customers choose where their Prime Intellect compute runs: GPU availability can be filtered by region (for example United States, Canada or Europe West), by data center and by provider cloud before provisioning, across more than fifty providers, and its training and evaluation stack (prime-rl and verifiers) is open source and runs on the customer's own hardware. Hosted Training and Hosted Evaluations are the exception: they run on infrastructure Prime manages, with no stated region.

Sourcedocs.primeintellect.ai/api-reference/check-gpu-availabilityread 2026-09-22

The Environments Hub is a community registry of ready-to-run environments for training and evaluation, spanning math, coding, games, search and tool use, and multimodal tasks, plus validated benchmark implementations, and Lab guides give training recipes for code generation, tool use and search agents. These are ready-made starting points for improving an agent. Environments are tasks and scoring for research, though, not production workflows or role specific agents that do work when selected.

Sourcedocs.primeintellect.ai/tutorials-environments/environmentsread 2026-09-22

Work on Prime Intellect starts when a person or the customer's own code calls it: hosted evaluations and training runs launch from the dashboard or CLI, and sandboxes and instances through the API or SDK. No schedule, event subscription or other seam that starts an agent's work without a caller is documented; cluster alerting notifies people about hardware health, not agents.

Sourcedocs.primeintellect.ai/tutorials-environments/hosted-evaluationsread 2026-09-22

Model choice is the customer's throughout: Prime Inference is an OpenAI-compatible API that routes requests to a range of model providers, evaluations run against any model on it or any OpenAI-compatible endpoint the customer brings, Hosted Training lets the customer pick among open-weights models (Qwen, Llama, NVIDIA Nemotron, gpt-oss and others) with per-model token prices, and trained LoRA adapters deploy behind an OpenAI-compatible endpoint.

Sourcedocs.primeintellect.ai/inference/overviewread 2026-09-22

Outside code reaches Prime Intellect through a documented REST API (instances, availability, disks, clusters, sandboxes and chat completions) with scoped API keys, the prime CLI, a Python SDK for sandbox lifecycles, and its open source verifiers and prime-rl libraries, with which customers build and publish their own environments to the Environments Hub. The API, CLI and SDK are stable and documented. No MCP server is published.

Sourcedocs.primeintellect.ai/llms.txtread 2026-09-22

Evaluation of the customer's agent is a core product: an environment packages a dataset of tasks, the harness the model runs in (tools, sandboxes, multi-turn interaction) and a rubric of reward functions and metrics that scores it, and Hosted Evaluations run that environment against Prime Inference or any OpenAI-compatible endpoint, from the dashboard or CLI, storing each run with its logs and results in Prime Evals.

The Environments Hub adds validated benchmarks such as AIME, MATH-500 and Humanity's Last Exam. Agents can be tested on datasets before production and scored the same way across runs.

Sourcedocs.primeintellect.ai/tutorials-environments/hosted-evaluationsread 2026-09-22

Browser use exists only inside training environments. The verifiers library's BrowserEnv lets an agent navigate and act on web pages in a DOM mode through Stagehand or a vision-based computer use mode, but the browser itself runs on Browserbase, a separate service with its own account and billing, sold beside the platform rather than part of it. Prime Sandboxes are VMs for running code, not a documented browser or desktop the agent operates. Prime Intellect hosts nothing for operating web interfaces or desktop environments where APIs are absent.

Sourcedocs.primeintellect.ai/guides/browser-environmentsread 2026-09-22