Translate plain English prompts into visual graphs to truly understand what your prompt is doing. Build single agents or full multi-agent systems with master orchestrators and linked sub-agents. Catch hidden errors, empty paths, warnings, and dangerous actions before you deploy.
Workflow
From a plain English description to an analyzed, executable agent graph — in four steps.
Type a natural language prompt describing the logic you want to visualize. Translating it to a graph helps you truly understand what your prompt is doing — catching empty paths, warnings, and dangerous actions.
Gemini 3 Flash reads your prompt and builds a complete node graph — typed, Dagre-positioned, and connected.
Drag nodes visually or type commands in the chat panel — "Add a retry loop around Search Tool".
AI analyzes conflicts, auto-fixes issues, then simulates your agent in Simulation Studio — two modes, data flow tracking, and rich error reporting.
Visual Graph Editor
The interactive canvas uses intelligent auto-layout via Dagre — nodes are automatically positioned in a clean top-down flow. Annotation nodes (guards, rules, tools, memory) are styled distinctly with dashed borders. Drag nodes to reposition, connect them with typed edges, and use the minimap to navigate large graphs.
Prompt Architecture
Paste your agent's system prompt or natural language logic. Verto maps the instructions into a structural graph, allowing prompt engineers to audit how the AI interprets boundaries, tools, and transitions. See the "Mental Model" of your agent instantly.
Verto translates your text instructions into a verifiable architecture. Audit the branching logic, verify safety guards are in-path, and ensure no orphaned edge cases.
LLM Output Formats
Verto supports three output formats from the LLM — each producing the same graph but optimized for different trade-offs between readability, token efficiency, and generation speed.
Full verbose JSON with all fields explicitly spelled out. Maximum readability, ideal for debugging and manual inspection.
Indentation-based format with no braces or brackets. Excellent readability, slightly more tokens than JSON.
Array tuples with type codes and integer IDs. Decoder expands to full graph. Up to 66% fewer tokens.
Graph Chat
The Graph Chat panel sends your instruction plus the full graph context to Gemini. It interprets commands like "add a retry loop around Search Tool" and generates precise graph mutations — adding, removing, or updating nodes and edges. Every edit shows stats: nodes added, removed, or updated.
Ctrl+Enter to send
The Graph Chat panel uses Gemini to understand your instructions in context. It reads your entire graph, figures out the right mutation, and applies it — adding, removing, or updating nodes and edges.
Bidirectional Re-sync
Re-sync reconstructs a prompt from any graph state — no AI call needed. Use the dropdowns to pick which two sources to compare: your original prompt, any saved version snapshot, or the live graph. Word-level diffs and a similarity score quantify exactly how much things have drifted.
| 1 | You are a customer support agent. |
| 2 | Classify incoming user intents accurately. |
| 3 | Route billing queries to human agents. |
| 4 | - Store interaction history to memory after each turn. |
| - Applies safety filter before all responses. | |
| - Blocks inappropriate content automatically. | |
| 5 | Escalate unresolved cases after 3 failed attempts. |
| 1 | You are a customer support agent. |
| 2 | Classify incoming user intents accurately. |
| 3 | Route billing queries to support agents. |
| - Store interaction history to memory. | |
| 4 | - Applies safety filter before all responses. |
| 5 | - Blocks inappropriate content automatically. |
| 6 | Escalate unresolved cases after 3 failed attempts. |
Use the dropdowns to choose any two sources to diff: original prompt, any saved version snapshot, or the live graph. Re-sync walks the graph using stored source positions, node order, and logic snippets to reconstruct a faithful prompt — no AI call. The Jaccard word-overlap score gives you a single number: how far apart the two sources are.
Visual graph edits are intuitive but can silently lose intent. Compare any two snapshots — including branched versions — to quantify exactly what changed between any two points in your agent's history.
Branch Version Control
Every node add, delete, and edit auto-commits a snapshot. Versions branch like git — edits on v1 create v1.1, v1.2; a fresh start creates v2. The SVG branch tree shows exactly which version each commit descended from.
Unlike a flat version list, Verto tracks which snapshot each commit descended from. If you restore v1 and make edits, those edits branch off v1 — not from the latest. The visual tree makes this lineage immediately visible.
AI Conflict Detection + DAG Validation
Two validation layers protect your graph. The DAG Validation Engine runs 15 deterministic structural rules instantly — checking for cycles, self-loops, disconnected components, missing source/sink nodes, and topological sortability. The AI Conflict Analyzer uses Gemini for LLM-powered risk detection, guard bypass analysis, and semantic reasoning — finding unguarded dangerous actions, logic contradictions, range gaps, and prompt/graph divergence. Click any detected action to navigate directly to its node on the canvas. Filter by severity and auto-fix issues with one click.
"Send Notification" has a GUARD node, but there exists an alternative path that reaches this action without passing through the guard.
15 deterministic DAG rules run instantly. The AI analyzer uses Gemini to detect risky actions, classify their permissions, and find guard bypass paths — plus semantic reasoning for issues simple linting would miss. Click any permission to pan directly to its node. Enable DAG-aware AI generation in Settings > Graph Rules to have Gemini actively avoid violations during creation and editing.
Validate immediately after AI generation
Reject edits that create critical DAG violations
Validate before executing an agent
Inject DAG rules into AI prompts (~1.5-2x tokens)
For fixable issues, the AI generates a precise graph mutation — adding missing edges, inserting guard nodes, or updating stale values — applied with one click.
Simulation Studio
A unified simulation environment with two modes, pre-flight validation, data flow tracking, and rich error reporting — all in a 3-column layout.
"My invoice shows the wrong amount for February"
{ "intent": "billing_dispute", "confidence": 0.94 }1 LLM call total
Makes a single Gemini call to generate realistic sample data for every node in your graph, then walks through them instantly. Fast, cheap, and great for validating structure.
1 LLM call per executable node
Runs your agent like real execution — each TASK, ACTION, DECISION, and GUARD node gets its own Gemini call. Produces realistic per-node outputs and real decision evaluations.
Comprehensive Taxonomy
Every part of an AI agent workflow has a purpose-built node — from personas and triggers to memory stores, safety guards, and logging outputs.
Quick Start
Five pre-built agent templates get you to a working graph instantly — no prompt needed.
Full Feature Set
From first draft to production-ready agent — Verto covers the entire workflow.
Multi-Provider AI
Plug in your own API key for any supported provider. Default is Gemini 3 Flash.
In Development
Features currently in development. Expect rough edges and breaking changes.
Translate your prompt into a graph to truly understand what it's doing. Expose hidden errors, catch dangerous actions, and identify empty execution paths before they break your application.
Open Agent Architect →No account required · Runs entirely in your browser · Free to use
I'm constantly improving Verto Agent Architect. If you encounter any issues, have suggestions for new features, or just want to chat about agent design, feel free to reach out.
project.verto@proton.me