Back to Devoured - July 14, 2026
DEVOURED
Today, we are releasing verifiers v1

Today, we are releasing verifiers v1

AI llmagents Prime Intellect
Prime Intellect's verifiers v1 platform standardizes environment stacks for agentic reinforcement learning, enabling long-horizon task training without context window limitations.
What: Prime Intellect's verifiers v1 decomposes environments into task sets, harnesses, and runtimes, using a managed interception server to store traces as message DAGs. This approach reduces trace memory complexity from O(n²) to O(n) and integrates directly with their prime-rl trainer, achieving 1,000 steps in two days on 6 H200 nodes.
Why it matters: By standardizing how agents interact with terminal environments and record state transitions, developers can now scale RL training for complex agentic workflows like coding or computer use without building custom infrastructure for every new model.
Deep dive
  • Decomposed stack: Splits training environments into three distinct layers (task, harness, runtime) for modularity.
  • Message DAGs: Stores rollout data as directed acyclic graphs to treat one trace as multiple training samples.
  • Efficiency: Reduces storage and processing overhead for long-horizon agents.
  • Scalability: Enables high-throughput training on limited GPU hardware (e.g., 6 H200s).
  • Integration: Plugs directly into the prime-rl framework.
Decoder
  • Agentic RL: A reinforcement learning approach where an AI agent learns to navigate multi-step environments to achieve a specific goal.
  • DAG (Directed Acyclic Graph): A data structure where nodes are connected in a way that there are no loops, perfect for representing non-linear process traces.
  • Harness: A testing framework that provides a consistent interface for an agent to interact with different environments.
Original article

Today, we are releasing verifiers v1 — an overhaul of our environment stack for the modern era of agentic RL and evals.

We decompose environments into a taskset, a harness, and a runtime.

Run complex agentic tasks like coding and computer use at scale, in any harness.

The decomposition allows you to:

  • Create tasksets, either verifiers-native or by re-using other frameworks such as Harbor
  • Evaluate and train in any harness (e.g. Codex, mini-SWE-agent) or bring your own
  • Run them in different runtimes, e.g. subprocesses, Docker or sandboxes

For example, run Nemotron 3 Ultra on Terminal-Bench 2 under Codex by authoring a simple config and launch it with our CLI.

The central piece is the verifiers-managed interception server which proxies the requests between the harness and the inference server.

It records traces on the fly, which allows for training and rewriting (e.g. to mitigate reward hacks).

These traces are now message DAGs: every message is stored exactly once.

Now, trace sizes are O(n) in turns instead of O(n²), which makes long horizon agentic rollouts feasible, especially for router replay and multimodal data.

Rollouts aren't linear. Compaction and subagents create branches in the graph — and every root→leaf branch is a contiguous, trainable sample.

One trace, N training samples. Long-horizon training past the context window, natively.

verifiers v1 plugs straight into prime-rl for training.

We have been using v1 internally for all our production runs.

In this run, we train GLM-4.5-Air on ScaleSWE tasks with under-4-minute steps and 35-turn rollouts, completing 1K steps in 2 days on just 6 H200 nodes.

Read the full verifiers v1 deep dive: primeintellect.ai/blog/verifiers…

Explore all 41 articles in today's digest

Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi