May 2026

🤖 AI Agent Components in flyte.ai

The new flyte.ai submodule ships pre-built components for agentic applications:

• flyte.ai.agents.CodeModeAgent wraps the orchestrator sandbox with a customizable litellm-compatible LLM generator.
• flyte.ai.AgentChatAppEnvironment serves chat-shaped agents.

End-to-end examples cover vLLM serving, Claude Code, and token-level batching with TokenBatcher inside a FastAPI app.

🔧 Hierarchical Settings

A new Settings hierarchy spans the SDK, control plane, and Console.

• The interactive flyte settings edit CLI command views and edits hierarchical settings.
• Tasks created without explicit CPU, GPU, memory, or storage values inherit org- and domain-scoped defaults.
• A Console Settings page exposes the same hierarchy, with org-level controls including configurable retention for user-built images.

✨ Interactive Triggers

Triggers grow from a list view into a fully interactive surface.

In the SDK:

• Pass arbitrary input through a trigger with custom_context.
• Attach multiple triggers or links in a single @env.task decorator.
• Use first-class sleep tasks for explicit delays.

In the Console:

• Create a trigger directly from a task’s details page.
• Bulk-delete triggers, with confirmation.
• View the full trigger spec on the details page.

🔨 Non-Root Images by Default

Default built images now create a non-root flyte user and switch to it at the end of the build. Files added with with_source_folder, with_source_file, and with_code_bundle are copied with the right ownership automatically.

✨ Actionable CLI Error Messages

Common failures now surface as actionable error messages: missing Docker, missing kubectl, missing source files, unpicklable deployments, and user-module load failures. Uploads automatically retry on transient network errors.

🔨 Image Building Enhancements

• External images. A new reference architecture covers bringing your own externally built images into Flyte v2 workflows.
• --builder. A top-level --builder flag on flyte and flyte build resolves from config when not provided.
• Custom buildx builder. FLYTE_DOCKER_BUILDKIT_BUILDER_NAME selects a custom buildx builder.
• Dependency quarantine. The exclude-newer uv option quarantines new dependency releases for five days to protect builds from upstream regressions.
• Cache aliases. Cache behavior accepts common aliases like enable and off.

✨ Console Enhancements

• Task details gain a Versions tab for side-by-side comparison, the entry file path, descriptions, and sub-action totals.
• App details pages show parameters, descriptions, and a delete action, with a Kubernetes service account selector in the launch form.
• Run logs get a full-screen button and an actor filter.
• Policies and roles get confirm-delete dialogs.
• Logout. The Console adds a logout option.

💻 TUI and Shell Task Improvements

• Trace output is now readable for long-running runs in the TUI.
• Shell tasks gain a defaults parameter, plus rendering fixes.
• Directories can be passed alongside files into sandbox-based agents.
• Long-form descriptions render in flyte --help.

✨ New Plugins and Examples

New plugins land for Hugging Face datasets, Hydra, omegaconf, and papermill.

New examples cover conditional caching, custom-context triggers, an autoresearch loop, genomics helper utilities, and a Vue app alongside the existing FastAPI, Streamlit, and Gradio examples.

🔧 User-Facing Logger

flyte.logger is now a user-facing logger, separated from the internal SDK logger, with an INFO default and no prefix.

🔧 Keyring Opt-Out

The new disable_keyring config option skips storing and retrieving tokens from the system keyring. Use it in CI and headless environments.

⚙️ Cluster-Aware Data Access

Workflows running against a specific data plane can resolve their data proxy directly. Authentication continues to land on the control plane even when you use the data plane endpoint directly.

⚙️ OpenShift Support for Self-Managed Deployments

Self-managed deployments gain platform-level OpenShift support:

• Restricted-module installs.
• AWS EBS CSI driver.
• ECR pull secrets.
• AWS pod identity webhook.
• Post-install helper tooling.

The Helm chart also adds:

• external-secrets v1 API support.
• A configurable worker-pod security context.
• Connector logs configuration.
• A per-deployment network policy toggle.