docs(agents): define the persistent collaborator vision

[loganj]

🤖

Summary

People should be able to work with one recognizable agent across conversations and machines, without tracking its sessions or reading its worker chatter. This draft proposes that experience: an agent that remembers relevant context, follows work it is authorized to see, and coordinates its own tasks across multiple machines. It proposes a vision; it does not ship new behavior.

The agent's conversation carries decisions, meaningful progress, and results. People can deliberately inspect its internal work—what it delegated, where work is running, and what happened—without routine coordination filling their conversation or unread counts. Permission requests and failures that need a person still reach them.

Start with VISION_AGENT_COLLABORATION.md for the proposed collaborator model. VISION_AGENT.md retains its existing purpose and runtime semantics for the buzz-agent coding runtime and buzz-dev-mcp tool server, with only a cross-reference added. The main vision links to the new document; the remote-agent and activity-feed visions address concurrent machines and deliberate inspection within their existing responsibilities. No code, configuration, or AGENTS.md changes are included.

What changes in the vision

Buzz already distinguishes a reusable persona—the model and instructions describing an agent—from its signed identity and the software running it. The existing remote-agent vision already promises that the same identity can return on replacement compute.

Today, device-held keys and settings, deployment-provider boundaries, and separate conversation sessions still shape continuity and control. Shared durable memory does not keep every live session's context synchronized. In the Janet/pic worker experiments examined for this proposal, internal worker messages use ordinary conversation threads and inherit their unread behavior, despite separate inspection tools; this is not evidence of shipped Buzz worker behavior.

The proposal makes the agent responsible for continuity and work coordination across machines, with internal work available for deliberate inspection. Concurrent machines extend the prior vision of one running instance while preserving duplicate-deploy protection within each provider's deployment scope. Community authorization still governs access and action. How keys reach machines, context stays current, and tasks get assigned remains for later design.

Related issue

Closest adjacent PR: #5419 (multi-host agent identity). Related discussions: #4174 (shared remote execution) and #5282 (persistent hosts). #7313 is an adjacent room-model vision draft; this proposal does not adopt its room architecture or supersede these discussions.

Buzz discussion (open in Buzz): buzz://message?channel=03856dc7-88aa-44ba-8fa5-49e9ce2950d7&id=22b825978258378d4863fc0b55e89a41190c5525d7607fe92ed6acf1adc9b2be.

Testing

Documentation-only: reviewed the complete diff, relative links and anchors, Markdown structure, and file-size policy. The earlier full just ci attempt timed out during dependency preparation; full local CI has not passed. No runtime behavior was changed or exercised.

[github-actions]

🔐 Codex Security Review

Status: review required for the current range.

The current range is cd02b693aae86444b0b5636474a3d202c6fc8f26...3cfdcbef568cc54117748dd96d4c343c4c3a27bc.
A new review must complete for this exact range. When manual authorization
is required, a Block organization member must comment exactly
@buzz-security-review 3cfdcbef568cc54117748dd96d4c343c4c3a27bc to authorize a new review.
Any previous review applies only to its recorded range.