In the previous release, I introduced Pixel Council - a visual, canvas-based multi-agent orchestration workspace where autonomous AI agents collaborate along pipeline connections. Today, we're shipping a major update focused on collaborative transparency and real-time Human-in-the-Loop (HITL) steering: Watch-Along Sessions and Instant Spectator Feedback.
What’s New in This Release
1. Watch-Along Sessions via Shared URL
You can now share your active agent canvas with teammates, reviewers, or stakeholders using a single lightweight URL.
- State Encoding without Database Friction: The workspace topology, agent configurations, and memory states are serialized into URL payloads, allowing observers to load and inspect the exact council state instantly.
- Spectator Mode: Watchers get full visual access to agent characters, active pipeline signals, execution logs, and output streams without accidentally mutating the canvas geometry.
2. Human-in-the-Loop (HITL) Spectator Bar
Observation is only half the story. The new Spectator Bar enables reviewers to intervene and guide agents dynamically while they execute tasks:
- Targeted Feedback: Direct suggestions, revisions, or praise to a specific agent or the entire council.
- Live Memory Injection: Feedback is instantly appended to the target agent’s working memory context and goal constraints.
- Immediate Re-evaluation: On subsequent execution ticks or pipeline triggers, agents naturally adapt their reasoning based on spectator input.
3. Dynamic Canvas Status Indicators
To make agent activity intuitive at a glance, pixel character nodes on the spatial canvas now feature color-coded real-time status badges:
- 🟠 Busy / Processing: Pulsing indicator showing active LLM generation.
- 🟢 Idle: Ready to receive signals or feedback.
- 🔵 Sleeping / Cooldown: In-between pipeline turns.
-
State Serialization:
exportEncodedState()encodes canvas nodes, links, and workspace state into a safe Base64 string parameter (?session=...&state=...). -
Context Augmentation: When a spectator submits feedback,
addSessionFeedback()appends the review text intoagent.memoryand appends a human constraint note toagent.goal. - Execution Integration: During the next pipeline turn, the agent prompt generator pulls the updated memory stack, ensuring human feedback strictly conditions the model's output.
Click Share & Watch in the top bar to copy your live session URL, send it to a teammate, and let them guide your pixel agents in real time!
Code & more: https://www.dailybuild.xyz/project/205-pixel-council
Top comments (1)
Serializing the canvas, agent configurations, and memory into a shareable URL makes a watch-along session feel immediate, while spectator mode prevents reviewers from accidentally rearranging the workspace. The more interesting product decision is letting feedback enter both
agent.memoryandagent.goal, so human guidance becomes part of the next execution rather than a comment that gets lost beside the run. For founders, the key tradeoff is trust: live intervention can make autonomous systems more useful, but only if the session clearly shows what changed, who changed it, and which downstream actions were affected.