Bounded Orchestration Nexus for Structured Agentic Integration

BONSAI

A Mixed-Initiative Workspacefor Human–AI Co‑Developmentof Visual Analytics Applications

BONSAI replaces unconstrained code generation with bounded orchestration across a four-layer stack—so VA systems stay modular, traceable, and under expert control.

Join the Waitlist

The Challenge

Unconstrained AI Breeds Opaque Monoliths

Fast prototyping is valuable—but “vibe coding” loses the causal thread. BONSAI is built around three non-negotiables:

  • R1 · Composability

    Hard boundaries between layers and services so teams can scale and reuse without entanglement.

  • R2 · Bounded Autonomy

    Agents run with explicit scopes, validated interfaces, and guardrails—automation that respects architecture.

  • R3 · Provenance

    Decisions and handoffs are first-class events: a legible trail from intent to deployment.

The Bonsai Core

Four Layers, One Discipline

The stack in one glance: what users see on top, how work is coordinated in the middle, and how compute is pooled and governed at the bottom.

L4 · Application

This is the face of your tool: charts, controls, and guided steps where analysts work, while the platform runs the pipelines behind the scenes.

L3 · Orchestration

Here multi-step workflows are stitched together, executed in a sensible order, and updated when inputs change—without you micromanaging every handoff.

L2 · Service

Small, focused pieces each do one job and plug in the same way everywhere, so teams can reuse work instead of rebuilding the same logic in silos.

L1 · Hardware

Office machines, cloud capacity, and heavy-duty clusters are treated as one pool, with clear rules so sensitive runs stay in the right environment.

BONSAI four-layer architecture: application, orchestration, microservices, and hardware scheduling
Layered Core Architecture

The Workspace

Plan · Design · Monitor · Review

A mixed-initiative loop: humans set goals and policies; the orchestrator decomposes work and delegates to specialized units.

Plan

Goals, constraints, brainstorming with feasibility signals, issue board, and configured ADUs.

Design

Service registry, typed pipeline editor, application view—new services become nodes without downtime.

Monitor

Agent map plus provenance with semantic zoom—every handoff is observable.

Review

Merge-gated progress, live preview, and issues filed straight from the running prototype.

BONSAI workspace lifecycle: Plan, Design, Monitor, and Review phases with representative UI for each stage
Plan · Design · Monitor · Review

Human–AI Interaction Workflow

Human–AI interaction workflow: iterative devolve cycle across Plan, Design, Monitor, and Review phases
Iterative Devolve CycleHow Humans and Agents Co-Develop VA Systems Across the Four Workspace Phases.

Monitoring Layer

In Monitor, a provenance graph and a spatial agent map make progress and dependencies legible—zoom from the big picture into individual tasks.

Provenance tree: timeline graph with swimlanes, semantic zoom detail, related-node highlighting on hover, and rich tooltips
Provenance TreeHierarchy, Dependencies, and Activity over Time—with Semantic Zoom and On-Demand Context.
Agent map: isometric 3D grid of workspace rooms with status colors, overview mode, and drill-down to ADU desks
Agent MapOverview of Rooms and Agents; Pair with the Provenance View to Relate Spatial Layout to Task History.

AI Integration

Structured Agency—Not Parallel Chaos

Three tiers mirror how serious software teams actually ship.

  1. The Nexus

    Dialogue → typed intents, backlog, policy gates, and service mining before spawning new work.

  2. Squad Leads

    Split issues, acceptance criteria, and a tight plan → clarify → declare → implement loop with rework queues.

  3. AI Development Units

    Specialized agents, one layer each, isolated branches, scoped task packs—no shared working context.

Mission Control and Nexus orchestrating two squad leads (User Registration, Dashboard Analytics), each with Frontend, Backend, and specialist ADUs, merging work into feature branches and main.
Nexus delegates to squad leads; ADUs work in isolation on scoped branches before merge to main.
  • Service LevelSkill files encode conventions and guardrails per layer.
  • Orchestration LevelMCP exposes design-time ops as bounded calls.
  • Application LevelVACP embeds MCP inside the app under construction.

Validation

System Validation & Use Cases

We evaluate BONSAI with two longitudinal case studies: real visual analytics work spans weeks or months, not a single controlled session. Semantic Color Mapping stresses catalog reuse—Nexus composed microservices into a DAG, CType exposed a field-level schema mismatch between stages, an adapter fixed it, and two ADU cycles produced an end-to-end pipeline without new services. PODIUM had no catalog fit, so the system was split into independent CType-aligned services built from scratch: a modular teaching prototype that underscores how costly interface changes become once OpenAPI contracts are fixed downstream.

Semantic Color Mapping

Semantic color mapping pipeline: flow inputs through aggregation, vector representations, unit of analysis, projection, and color map to the final web application.

Mennatallah El-Assady, Rebecca Kehlbeck, Yannick Metz, Udo Schlegel, Rita Sevastjanova, Fabian Sperrle, Thilo Spinner. Semantic Color Mapping: A Pipeline for Assigning Meaningful Colors to Text. 2022. IEEE Workshop on Visualization Guidelines in Research, Design, and Education (VisGuides). DOI

PODIUM

PODIUM live app in BONSAI Review: tabular ranking dashboard with in-context issue reporting.

Emily Wall, Subhajit Das, Ravish Chawla, Bharath Kalidindi, Eli T. Brown, Alex Endert. Podium: Ranking Data Using Mixed-Initiative Visual Analytics. 2018. IEEE Transactions on Visualization and Computer Graphics, vol. 24, no. 1, pp. 288–297. DOI

  • Contract-First
  • Reuse-First
  • Merge-Gated
  • Full Provenance
  • Bounded Agents

Gallery

Diagrams and Screenshots

Extra visuals for the concepts on this page—architecture, workflow, and validation examples.

WorkspaceMiddleware portal: typed pipeline editor and live application preview.
ArchitectureFour Layers from User-Facing Apps down to Pooled Compute.
Human–AI WorkflowHow People and Agents Move through Plan, Design, Monitor, and Review.
Orchestration (SCM)Example of Wiring Services for the Semantic Color Mapping Scenario.
Provenance TreeTimeline and Hierarchy with Semantic Zoom, Hover Highlighting, and Tooltips for Each Handoff.
Agent Map3D Overview of Rooms and ADU Desks; Drill Down from the Map into Task-Level Provenance.
Map DetailCloser Look at a Region of the Workflow Map.
PODIUM Live AppIn-Context Feedback on a Running Prototype.
Middleware & AgentsHow Middleware Connects Tiers of Automation.
Semantic Color Mapping FlowStages from Text to UI: Aggregation, Embeddings, Analysis Units, Projection, and Color Map—Services Sit at Explicit Boundaries.
PODIUM PipelinePipeline Layout for the PODIUM Example.

Build Your Next VA App with BONSAI

Sustainable multi-agent development needs structure and oversight—like the tree it’s named for. Join the waitlist for early access.