The Nine Capabilities
The nine things AgentCanvas does that, together, define it as a platform โ each with its own implementation page covering the mechanism, key files, and current status.
1Graph-Expressible Agents
DAG workflows, loops (rollout or agent), and compositions of both run on the same canvas and executor.
2Customizable Node System
Subclass BaseCanvasNode, register in WorkspaceComponentRegistry โ new node types appear in the workspace automatically.
3Isolated Runtime Environments
Server-mode nodesets run in their own subprocess and Python env; the framework talks HTTP to them.
4Nested Graph System
Composite nodes contain subgraphs; flatten.py recursively expands them before execution.
5Graph Execution Engine
Fire-on-input ready-queue scheduler โ no form-specific code paths for DAG vs. loop vs. mixed.
6Execution Logs & Live Views
Two-layer JSONL logs of every node firing; live viewer windows on the canvas โ query a run afterwards, watch it as it happens.
7Batch Evaluation & Job Queue
Submit a graph as an eval run; GPU-budget-gated subprocess execution with per-episode persistence; one queue shared across all Claude / canvas sessions hitting the same backend.
8Visual Canvas Editor
React Flow editor for authoring graphs, configuring nodes from auto-rendered forms, and watching every port update live over one WebSocket โ root canvas and composite drill-down share the same component.
9Graph SDK
Build and run the same graphs as a Python library โ import Graph, add/connect nodes, run and batch-eval in-process, or compile a graph back into a builder script. Same GraphDefinition, fully reversible with canvas + JSON.
10Coding-Agent Harness
Stock coding agents โ Claude Agent SDK, mini-swe-agent, Codex CLI โ as VLN agents over one byte-equivalent tool surface, scored by habitat's own ruler
ยง Related
- Blueprint โ the conceptual statement of the capability set
- Design Docs โ deeper internal references referenced by these pages
- Decisions (ADRs) โ the why behind the wiring