LoginSign Up
← Back to all sparks
W

Windmill

INFRA · APIS
Velocity7.5

Open-source developer platform to build internal tools and workflows from scripts

www.windmill.dev ↗

Windmill is building a full AI-native workflow platform — guest app access, agent evals, dbt runtime, and AI cost tracking shipped in rapid succession.

workflow-automationai-agentsdevtoolsdbtenterpriseobservability
◆Current state
Windmill is an open-source workflow automation and developer platform that has been adding AI capabilities at an accelerating rate. The last month brought AI agent evaluations (with dataset-based scoring and run comparison), AI usage and cost tracking per user and model, a Plan mode for AI sessions (read-only investigation posture), version history for both AI artifacts and resource values, and nested filter groups for Kafka and WebSocket triggers. The platform now spans workflow orchestration, AI-assisted script writing, and a native dbt runtime.
◆Where it's heading
The pattern across recent releases is clear: Windmill is building the operational infrastructure around AI agents — not just the ability to run them, but the ability to measure, control, audit, and govern them. Agent evals add a testing layer; Plan mode adds a safety posture; cost tracking adds financial visibility; version history adds rollback capability. The guest app access model also introduces a new monetization dimension (4:1 guest-to-seat ratio above 100 free guests). Windmill is positioning as the AI workflow platform for teams that need enterprise-grade controls.
◆Prediction
The eval framework (currently beta) will likely gain deeper integration with CI/CD and deployment flows. Guest access metering signals that Windmill is building toward consumption-based pricing for external-facing AI apps. Expect tighter integration between the agent evals system and the existing flow editor deployment pipeline.

◆Recent moves

  1. 8d ago

    Guest access for apps

    ⚡ SPARK

    Guest app access introduces a fourth access tier between members-only and public, backed by a new metering model (100 free guest emails/30 days, then 4:1 guest-to-seat above that on EE). This changes the economics of sharing Windmill apps externally and opens Windmill to client-facing and partner-facing deployment patterns that were previously impractical.

    View source ↗
  2. 20d ago

    AI agent evals (beta)

    ⚡ SPARK

    AI agent evals bring structured measurement to Windmill's AI agents: curate test cases, score with AI judges or code scorers, compare runs cell by cell. This is the first evaluation framework built directly into a workflow platform rather than bolted on externally, and it directly addresses the 'how do we know if this agent got worse' problem that currently has no good answer for most teams.

    View source ↗
  3. 20d ago

    AI usage and cost tracking

    AI usage and cost tracking per user, model, and provider turns AI spend from invisible to visible — workspace admins get aggregate dashboards, members see their own costs. The OpenRouter exact-cost path (vs. the estimated-cost path for other providers) shows Windmill is building first-class integrations with model routers, not just raw APIs.

    View source ↗
  4. 23d ago

    Application default credentials for GCP Pub/Sub triggers

    GCP Pub/Sub triggers can now use application default credentials (Workload Identity, metadata server, GOOGLE_APPLICATION_CREDENTIALS) instead of a stored gcloud resource, removing a credential management requirement for GKE-deployed Windmill instances. This is a meaningful improvement for teams running Windmill in GCP's managed identity model.

    View source ↗
  5. 1mo ago

    Plan mode for AI sessions

    Plan mode gives AI sessions a read-only posture where the assistant can investigate but cannot execute, deploy, or write — until the user approves a plan. This is a concrete implementation of the 'human in the loop' principle for agentic workflows and addresses a real concern about AI agents taking irreversible actions in production environments.

    View source ↗
  6. 1mo ago

    Nested filter groups and dotted paths in trigger filters

    Nested filter groups (any_of/all_of/none_of, up to three levels) and dotted-path field access in Kafka and WebSocket triggers remove a fundamental expressiveness limitation that required flat AND/OR logic. Teams filtering on complex event schemas — common in event-driven architectures — can now express compound conditions without pre-processing events upstream.

    View source ↗