Detailed Analysis
A developer has built a novel interface for Claude Code that reimagines multi-agent orchestration as a spatial, visual experience rather than a series of terminal windows or chat threads. The tool imports existing FigJam boards, preserving colors, sticky notes, arrows, and organizational sections, and lets users spawn Claude Code agents directly onto the canvas as visible objects. Each agent appears as a small terminal with its own cursor color, and critically, it receives context of the entire board, every sticky note, zone, and connection, rather than being fed a narrow prompt. This is a significant departure from how most people currently interact with coding agents: instead of copy-pasting context into a chat window, the agent is embedded in the same visual workspace where planning already happens.
The most compelling demonstration involves auditing a real SaaS company's messy FigJam board containing thousands of planned feature stickies. When asked whether everything in a specific zone was implemented, the agent visibly scanned the relevant stickies (shown via a glowing cursor tracing across cards), cross-referenced the company's GitHub repository and open pull requests, and returned a precise answer: seven of nine features implemented, with the two gaps pinned directly onto the board alongside links to relevant PRs. This illustrates a broader shift in how AI agents can be useful: not just generating code from a prompt, but reconciling stated plans against actual technical reality and leaving that reconciliation as a persistent, inspectable artifact in the same workspace where the plan lives.
Several design choices reveal where this kind of tooling is heading. Agents claim a "desk" on the canvas where connected tools like GitHub or Linear appear as tiles, giving at-a-glance visibility into what each agent is doing without opening logs. Sessions can be forked mid-task via copy-paste, cloning the full transcript and letting it diverge, an idea that hints at branching, parallel exploration as a first-class primitive for agentic work. A mission-control kanban view aggregates many agents by status, addressing a real pain point in multi-agent workflows: once you have five or ten agents running, tracking their state becomes its own management problem. Under the hood, the system uses real Claude Code sessions running in tmux, with hooks streaming tool calls and status changes to the canvas, plus a custom API letting agents read and write board content with access receipts, a detail that signals attention to auditability and trust, not just novelty.
This project sits within a broader wave of experimentation around agent orchestration UIs, as developers move past the assumption that a terminal or chat window is the natural home for coding agents. As agentic coding tools like Claude Code become more capable and more numerous in parallel use, the bottleneck shifts from model capability to human oversight: how do you supervise, compare, and steer many agents at once without losing context? Spatial interfaces borrowed from collaborative whiteboarding tools offer one answer, treating agents as first-class citizens alongside the planning artifacts they're meant to act on. The rough edges acknowledged, laggy performance on large boards and desktop-only availability, suggest this is very much a prototype, but it points toward a future where the boundary between "planning documents" and "executing agents" dissolves into a single, persistent, visual workspace.
Read original article →