Polyaxon Roadmap

Compare dense run views with charts, graphs, synchronized cursors, shared selections, and parameter and metric tooltips.

Implemented for the upcoming 2.18 release. Undeclared params become context values; strictParams enables undeclared-name checks. Declared input and output validation still applies.

Implemented for the upcoming 2.18 release. Components, operations, and compiled operations can omit version; existing versioned files remain valid. The field no longer selects a validation schema.

Debug live workloads with sandbox command execution, file access, and interactive terminals.

Connect local terminals, IDEs, and file-transfer tools to running services through Polyaxon's authenticated SSH tunnel.

Enable the tmux plugin to disconnect and return to the same shell from the CLI or UI. Attach to tmux explicitly over SSH to preserve your shell across SSH disconnects. Sessions retain their working directory, variables, and running programs while the container and tmux process remain alive.

Connect hooks and platform events to evaluations, notifications, and downstream jobs when runs change.

Manage built-in connections and integrations at the control plane level.

Bring events from Git providers, webhooks, and external systems into Polyaxon workflows.

Run bounded research campaigns that propose experiments, evaluate results, and retain evidence for each decision. Tune training configurations, agent prompts, or tool policies with explicit budgets and approvals. Track proposal branches, compare results against a baseline, and pause or resume across restarts. Configuration-only proposals come first, followed by isolated source experiments; numerical searches can delegate to native hyperparameter optimization.

Create personal access tokens with explicit permissions, service access, and expiration dates. Set a default token for new runs or select a specific token when submitting from the CLI or UI. Inspect token scopes and expiry, revoke access, and audit token creation and revocation. A staged rollout introduces scoped run credentials before retiring legacy token behavior as agents and CLIs are upgraded.

Connect a Kubernetes cluster through cluster credentials or a linked cloud account, with Polyaxon hosting the agent and managing workloads remotely. No Polyaxon CRDs, operator, agent, or streams pods are installed in your cluster, so idle clusters need no resident Polyaxon pods or nodes reserved for them. Bring your own S3, GCS, or Azure artifact storage, and see connection, permission, and storage issues in agent status.

Manage artifact storage alongside the workloads that produce it.

Access models through a gateway with fallbacks, load balancing, retries, custom guardrail hooks, and advanced evaluations.

Monitor agents, jobs, and clusters with built-in monitoring and observability.

Explicitly bind Polyaxon queues to existing Kueue, Volcano, or KAI queues for native Trainer workloads. Follow runs from platform dispatch through cluster admission and scheduling, with pending reasons and scheduler messages visible in Polyaxon. View external queue usage and quotas where available, alongside Polyaxon's own concurrency, quota, and access controls.

Targeting 2.18. Define jobs and services with less nesting, using container settings, commands, environment variables, and resources. Declare ports to select a service without an explicit kind field.

container:
  image: ghcr.io/astral-sh/uv:python3.12-trixie-slim

env:
  MODEL: Qwen/Qwen2.5-0.5B
  DATASET: stanfordnlp/imdb
  GPUS_PER_NODE: "1"

resources:
  cpu: "4..8"
  memory: "8Gi..16Gi"
  gpu: 1

cmd:
  - uv pip install --system trl
  - |
    trl sft \
      --model_name_or_path "$MODEL" \
      --dataset_name "$DATASET" \
      --num_processes "$GPUS_PER_NODE"

ports:
  - 8080

Targeting 2.18. Define simple runs directly with CLI options, or start a sandbox with --sandbox.

polyaxon run \
  --image ghcr.io/astral-sh/uv:python3.12-trixie-slim \
  --env MODEL=Qwen/Qwen2.5-0.5B \
  --env DATASET=stanfordnlp/imdb \
  --env GPUS_PER_NODE=1 \
  --cmd 'uv pip install --system trl' \
  --cmd 'trl sft --model_name_or_path "$MODEL" --dataset_name "$DATASET" --num_processes "$GPUS_PER_NODE"' \
  --gpu 1 \
  --cpu 4..8 \
  --memory 8Gi..16Gi

polyaxon run \
  --image busybox:1.36 \
  --cmd 'httpd -f -p 8080' \
  --port 8080

polyaxon run --sandbox \
  --image python:3.12 \
  --gpu 1 \
  --cpu 4..8 \
  --memory 8Gi..16Gi

Improve batch search with native TPE v2 and Gaussian process Bayesian optimization, benchmarked against Random and Grid search. Reuse imported evaluations without rerunning them and allocate trial budgets asynchronously with ASHA and TPE+ASHA. Versioned policies keep active studies on their original behavior, with one optimization engine for model training and agent evaluation.

Release compute when sandboxes are idle and wake them on access. Workspace files survive suspension; running processes and terminal sessions do not.

Reduce startup time for new runs with a pool of ready sandboxes. Used sandboxes never return to the pool.

Specify logic, hardware, and execution through composable programmatic primitives.

Test and debug agents with step-by-step execution, state inspection, and performance profiling in a sandbox environment.

Trace requests with OpenTelemetry-compatible observability. Capture inputs, outputs, and metadata for intermediate steps to identify bugs and unexpected behavior in LLM applications and agents.

Evaluate and monitor GenAI applications throughout their lifecycle to measure quality, compare changes, and detect regressions.

Manage and optimize prompts alongside application traces, evaluations, and model access in an integrated GenAI workflow.

Add comments to runs, models, artifacts, traces, spans, observations, and scores to share context and collaborate across teams.

Improve onboarding and core screens for new and nontechnical users, and improve UI performance for large dashboards.