Agent Builder
Design, test, and publish multi-step AI workflows on a visual canvas.
The Agent Builder lets you wire together AI agents, logic gates, data transforms, and tool calls into a graph that runs as a repeatable workflow. Draw nodes on a canvas, connect them with edges, configure each step, then deploy to production — all without writing backend code.
What you can build
| Use case | Nodes involved |
|---|---|
| Multi-agent pipelines | agent → classify → agent |
| Document review with human sign-off | agent → guardrails → user_approval → end |
| RAG + reasoning | file_search → agent → transform → end |
| Tool-augmented agents (MCP) | agent + mcp node → end |
| Dynamic delegation to specialized workflows | agent with the trigger_workflow tool → end |
| Scheduled or webhook-driven runs | Any workflow + a cron or webhook trigger |
Key Features
| 🎨 Visual canvas | Drag nodes from the palette, draw edges by clicking connectors |
| 🤖 AI Assist | Generate a full workflow from a text description, or get node suggestions for a workflow in progress |
| 🧩 Agent Skills | Package system prompts, tools, and knowledge into reusable versioned configurations attached to any agent node |
| 📋 Workflow lifecycle | Draft → Stage for Review → Published/Rejected (approve publishes directly — no separate approved state) |
| 🔢 Versioning | Snapshot any version, roll back instantly, diff two versions side-by-side |
| ▶️ Live run panel | Execute workflows and stream output in real time — includes elapsed-time counter and reconnect banner on connection drop |
| 🧪 Test cases | Save and replay named input fixtures to verify workflow behavior |
| 📤 Import / Export | Share workflow graphs as portable JSON bundles; skills, config, and the default skill all travel with the workflow |
| 🚀 Promote to catalog | One-click export + import — copy a published workflow into another catalog or space as a new draft (sandbox → uat → production) |
| 🔒 Role-based access | Separate roles for editors, reviewers, and publishers |
| 🏭 Distributed execution | Redis Streams–based worker pool for high-throughput, crash-resilient production deployments |
| ⚡ Event-driven workflows | Subscribe workflows to business events; fan-out across departments without coupling |
| 🔄 Saga orchestration | Coordinate multi-step processes across agents with automatic compensation on failure |
| 📊 Per-workflow analytics | Run history, success rate, avg latency, and token cost — accessible to developers without admin access |
| 🔁 while_loop node | Conditional routing with an iteration counter — repeat a subgraph path until a condition is met or a max iteration limit is reached |
| 🕵️ Traceability & audit log | Who-changed-what config trail plus a tool-call-by-tool-call execution trace, with agent-to-agent chain correlation for multi-agent runs |
Editor layout
┌────────────────────────────────────────────────────────────┐
│ Header: workflow name, status badge, lifecycle actions │
├──────────┬─────────────────────────────────┬──────────────┤
│ Palette │ Canvas │ Right panel │
│ (nodes) │ drag + drop, connect, zoom │ (props / │
│ │ │ run / AI) │
└──────────┴─────────────────────────────────┴──────────────┘
- Palette — four categories of nodes: Core, Tools, Logic, Data.
- Canvas — infinite scrollable space. Zoom with scroll wheel. Pan by dragging empty space.
- Right panel — switches between node properties, run output, evaluation results, and the AI Assist panel.
Quick start
- Go to Workspace > Agent Workflows and click + New Workflow.
- Give it a name, then click Open in Builder.
- Drag an
agentnode from the palette onto the canvas. - Click the node to open its properties. Set a system prompt or pick a skill.
- Draw an edge from the
agentconnector to anendnode. - Click Run (▶) in the top-right to test it with sample input.
- When ready, click Stage for Review → an admin approves → Publish.
Next steps
- Node Types — reference for every node and its config fields
- Agent Skills — Reusable agent configurations: system prompt, tools, knowledge, LLM defaults
- Workflow Lifecycle — Draft → Review → Publish flow
- Review Queue — admin queue for approving and rejecting staged workflows
- Inbox — human review tasks created by running workflows
- Versioning — Snapshots, rollback, and graph diff
- Export & Import — Portable JSON bundles: what travels, skill remapping, KB re-wiring; plus one-click Promote to catalog
- AI Assist — Generate workflows and get node suggestions from natural language
- Run Infrastructure — SSE streaming, pause/resume, distributed execution, crash recovery, rate limiting
- Analytics — Per-workflow run history, success rate, latency, and token cost
- Triggers — Cron, webhook, and event subscription triggers
- Built-in Tools — Tools any
agentnode can call, includingtrigger_workflowandemit_event - Event-Driven Workflows — Event Bus, subscriptions, external ingestion
- Saga Orchestration — Multi-step coordination with automatic compensation
- Unpredictable Process Patterns — Design patterns for processes with no fixed sequence
- Logging & Tracing — Structured logs, OpenTelemetry traces, and metrics for runs and nodes
- Traceability & Audit Log — Config change audit trail, tool-call execution trace, and agent-to-agent chain correlation
- API Builder — import and govern the REST APIs that back your agents' tool servers