[NeMo Gym] v0.7.0 Roadmap

[sephmard]

Release date: October 7, 2026

This is a summary snapshot of the planned work targeted for the NeMo Gym v0.7.0 release. Updates will be managed in the GitHub Project.

1. Runtime & Agent Infrastructure

Core platform for running agents, evaluations, and workflows.

• Improve the environment onboarding experience with scaffolding tools, clearer conventions, and guided workflows. [epic] Ease of use for creating custom environments  #1520
• Reduce coupling between specific agent harnesses and specific benchmarks so harnesses become swappable across benchmarks rather than hardcoded per benchmark. Decouple agent harnesses from benchmarks #2950, [epic] composable architecture consistency (swappable agents) #1866
• Split the agent server's two conflated responsibilities — rollout orchestration (/run) and the agent loop (/responses) — into separate, independently swappable concerns. decouple rollout orchestration and agent harnesses #2858
• Establish canonical patterns for how agents and resources servers use the shared Sandbox API, closing gaps in the current sandbox documentation. epic: canonical sandbox patterns #2763
• Submit evaluation jobs to SLURM clusters, Kubernetes, or heterogeneous node pools from a single shared config — with job tracking, result persistence, and reproducibility guarantees. [epic] orchestration: remote execution and job management for Gym evaluations #1640
• Improve data preparation with a unified dataset preparation workflow and standard source-format adapters so contributors onboard data once and the format works everywhere. epic: dataset preparation #1338
• Make rollout failures bounded, visible, and recoverable — report failures as policy-neutral facts so callers decide whether to retry, replace, mask, omit, or stop, rather than Gym silently changing eval/training semantics. Make rollout failures bounded, visible, and recoverable #2750
• Add partial rollout checkpointing so long-running executions don't lose all progress when a worker, node, or process fails. Partial rollout checkpointing for long-running Gym executions #3024
• Publish machine-readable rollout progress so external orchestrators can consume it reliably. Publish machine-readable rollout progress for external orchestrators #2957
• Define and enforce dependency pinning for server environments so a fresh run doesn't silently resolve a different third-party dependency set. [task] Define and enforce dependency pinning for server environments #2933

2. Training

Capabilities for creating training data, optimizing policies, and improving agents.

• Add a reference user_agent server that orchestrates multi-turn dialogue between a policy model and a configurable simulated user. feat: User agent simulation for multi-turn episodes #985
• Improve agent performance without retraining model weights through a pluggable family of context and prompt optimizers. [epic] agent-harness optimization: context/prompt optimizers #1549
• Teach aggregate metrics and reward profiling to read reward_components directly, eliminating duplicate top-level fields. Make metrics consume reward_components directly #2137

3. Evaluation & Benchmarks

Measuring, comparing, validating, and understanding agent performance.

• Add pairwise comparison, statistical significance testing, and grouped comparison across task slices so results drive defensible release decisions. [epic] Decision-grade analytics: pairwise comparison, release gating, and grouped comparison for Gym evals #1493

4. Ecosystem

Expanding supported environments, frameworks, and deployment targets.

• Add a built-in harness for LangChain Deep Agents. feat: Langchain Deep Agents Code Harness #2090
• Make Python object-oriented agent implementations — starting with NOOA — pluggable into Gym without requiring Gym-specific FastAPI routes or cookie handling. Object-oriented agents in Gym #2096
• Provide Dockerfiles and Kubernetes manifests so teams whose infrastructure requires Kubernetes (e.g. OpenShift) can run Gym. epic: Enable Kubernetes for the Gym for OpenShift #1269
• Ship CLI-native exporters. [epic] Exporting via Nemo Gym CLI #2901