Gitpulse
Merged
Size
XL
Change Breakdown
Refactor70%
Maintenance20%
Testing10%
#27925refactor(core): Remove multi-agent architecture entry point from AI workflow builder (no-changelog)

AI workflow builder simplified to single path

BU
burivuhster
·Apr 2, 2026

The AI workflow builder's dual-path architecture has been collapsed. Feature flags controlling two separate systems are gone — there's now one code path for all requests.

The n8n AI workflow builder previously supported two architectures: a legacy multi-agent system with separate discovery, builder, and responder phases; and a newer code builder path. A feature flag codeBuilder let operators toggle between them.

That toggle is now gone. The code builder is the only path forward.

The change removes roughly 3,600 lines across the codebase. Gone are the builder subgraph, multi-agent evaluation harness, process operations node, and dead configuration fields that existed only to support the now-removed architecture. The multi-agent graph still exists for plan mode delegation — discovery, planner, responder, and supervisor remain in place for handling plan requests, web fetch approvals, and plan modifications.

The experiment 071_coding_workflow_builder still exists but now only controls one thing: whether pin data gets generated. The code-no-pin-data variant is default; code-pin-data is the variant.

Telemetry remains unchanged — code_builder: true continues to be sent in events, and LangSmith traces still go to the code-workflow-builder project.

View Original GitHub Description

Summary

Removes the multi-agent system, making code builder the only path for the AI workflow builder. The codeBuilder feature flag is removed from the full stack — the experiment 071_coding_workflow_builder now only controls pin data behavior (code-no-pin-data as default, code-pin-data as variant).

What stays: The trimmed multi-agent graph (discovery + planner + responder + supervisor) is retained for plan mode delegation — plan requests, web fetch approvals, and plan modify/reject still route through it.

What's removed:

  • codeBuilder feature flag from frontend, API DTOs, CLI, and backend interfaces
  • Builder subgraph and its exclusive dependencies
  • Multi-agent specific evaluation harness
  • Legacy multi-agent routing path in WorkflowBuilderAgent.chat()
  • process_operations node (no-op without builder subgraph)
  • Dead config fields (instanceUrl, onGenerationSuccess, resourceLocatorCallback) from multi-agent graph
  • Obsolete tests for removed code paths

Telemetry: code_builder: true is always sent in the Builder replied to user message event and LangSmith traces always go to the code-workflow-builder project, preserving existing analytics.

Follow-up work (not in scope):

  • Strip supervisor agent (currently kept but only routes to discovery in plan mode)
  • Remove create*Tool functions from builder tools (display constants still used by session replay for past conversations)
  • Clean up StageLLMs interface (supervisor, parameterUpdater slots)

Related Linear tickets, Github issues, and Community forum posts

https://linear.app/n8n/issue/AI-2222/clean-up-feature-flags-based-on-recent-decisions

Review / Merge checklist

  • PR title and summary are descriptive. (conventions)
  • Docs updated or follow-up ticket created.
  • Tests included.
  • PR Labeled with release/backport (if the PR is an urgent fix that needs to be backported)
© 2026 · via Gitpulse