LoginSign Up
← Back to all sparks
D

Dify

AI-ASSISTANTS
Velocity6.3

Open-source platform for building, deploying, and operating LLM applications and agents.

dify.ai ↗

Dify is rebuilding itself around a sandboxed agent runtime, with the workflow builder as the legacy layer.

agent-runtimeworkflow-orchestrationhuman-in-the-loopskillsself-hostingtenant-isolation
◆Current state
Dify still ships as an LLM app platform — visual workflows, a knowledge base with vector retrieval, and self-hosted Docker deployment. But the last two release cycles have moved the center of gravity: a sandboxed Linux agent runtime, a Skill Editor for packaging reusable capabilities, and a Human Input node that lets a workflow pause for review. Between those, the releases are patch work: tenant isolation fixes, self-hosted SECRET_KEY hardening, and workflow-editor ergonomics.
◆Where it's heading
The arc from 1.13 to 1.16 is a conversion from graph-first to agent-first. HITL came first, making the workflow engine tolerant of pauses and external decisions; then the agent runtime arrived to fill those graphs with something that plans rather than follows edges. Dify Agent shipping as an explicit experiment — with a warning to expose it only to trusted users — signals the sandbox isolation is not yet production-grade, which is why the surrounding releases spend so much effort on tenant scoping and credential permissions.
◆Prediction
Expect Dify Agent to leave experimental status in a 1.17 or 1.18 release once the sandbox and credential-scoping work lands, with Skills becoming a shareable artifact alongside the existing app DSL export.

◆Recent moves

  1. 22d ago

    Release v1.16.1 - Bug Fixes and Security Enhancements

    A patch release that is mostly workflow-editor ergonomics: multi-select tool inputs, clicking a node_id in a run log to jump to that node on the canvas, and better keyboard navigation in the block selector. These are the fixes you ship when the graph builder has become the thing people debug rather than the thing they demo.

    View source ↗
  2. 1mo ago

    Dify Agent: a sandboxed shell agent you build from Skills

    ⚡ SPARK

    Dify Agent lands as a distinct product surface: a shell-based agent running in a Linux sandbox, built through a UI where you set a base prompt and upload Skills. It is the point where the agent runtime stops being infrastructure under the workflow engine and becomes something a user builds directly.

    View source ↗
  3. 3mo ago

    v1.14.2 - Security fixes, agent groundwork, workflow reliability, and deployment updates

    Security and reliability work laying groundwork for the agent runtime: tenant-scoped trace-config and file-preview endpoints, and builtin tool credentials restricted to workspace admins. The credential scoping matters more than it reads — an agent that can call tools inherits whatever the tool credential can reach.

    View source ↗
  4. 3mo ago

    v1.14.1 - Security hardening, workflow stability, and cleaner self-hosted deployments

    Housekeeping for self-hosted deployments: Docker installs no longer fall back to a published default SECRET_KEY, and internal /threads and /db-pool-stat endpoints are no longer unauthenticated. Real fixes, but nothing a user of the product notices.

    View source ↗
  5. 6mo ago

    Sandboxed agent runtime and a Skill Editor arrive in 1.14.0-rc1

    ⚡ SPARK

    The first release where the agent runtime and Skill Editor appear together — sandboxed execution, an Agent Mode, and reusable SOP blocks callable inline with @. Everything in the 1.16 line is a continuation of the direction set here.

    View source ↗
  6. 6mo ago

    1.13.0 - Human-in-the-Loop and Workflow Execution Upgrades

    ⚡ SPARK

    The Human Input node makes a workflow pausable, holding execution at a decision point for review and edit before continuing. In hindsight this is the prerequisite for everything that followed — an autonomous runtime is only deployable if the graph around it can stop and ask.

    View source ↗