Human Agents Workspace

The emerging category where AI agents are members of the workspace, not tools called from a terminal: shared rooms, identities, scoped access, and — the defining property — one shared audit trail where human and agent actions are first-class peers. The cluster’s earlier framing: paperclip (OSS) as the agent control plane (tickets → agent runs → review gates). Buzz adds the communication plane.

The components

• buzz (OSS) (35.1k★, Apache-2.0, Rust; @bibryam announcement captured) — Block’s self-hostable Slack-like workspace “where humans and AI agents share the same rooms”: agents get identities and scoped access; chat, git, workflows and approvals share one audit trail; every action is a signed Nostr event. The Nostr choice is the architectural bet — a signed-event log as the substrate means the audit trail is cryptographic, replayable, and transport-neutral from day one.
• paperclip (OSS) (84k★, MIT, TypeScript) — the other half of the pairing: org-wide agent management for tickets and workflows with review gates. Paperclip answers “which agents run what, gated how”; Buzz answers “where humans and agents talk and decide”. A complete workspace wants both planes wired to the same trail.

What makes it a category, not a feature

• Agent identity — an agent with a room presence, scoped permissions, and a signed action history is accountable in the way a service account never was (you can @ it, review it, revoke it)
• Shared audit trail — chat, git events, workflow runs, approvals in one ordered log; the compliance story stops being “grep three systems”
• Approvals as protocol — human-in-the-loop stops being a UI checkbox and becomes a signed event in the stream (the review-gate idea from paperclip, generalized)
• Self-hostable — both anchors are Apache/MIT and run on own metal; the PII redaction / Agent sandbox boundary stack composes: agents talk in the workspace, execute in sandboxes, egress through redaction

Open questions: does Nostr-as-audit-substrate generalize beyond Buzz, or is it a Block-scale eccentricity? Where does the harness (Hermes Agent) sit — as a room member via the same protocol? And does the control plane (paperclip) absorb the communication plane, or do they stay federated peers?

Related: Agent knowledge systems (typed-graph side: the work itself as a queryable base) · paperclip (OSS) · buzz (OSS) · Hermes Agent (harness side) · Agent sandbox (execution side) · PII redaction (egress side)

• Reclassification (2026-10-07): paperclip also qualifies as a Meta harness — the executive pole (org charts, budgets, goal alignment) vs Omnigent’s developer pole (sessions, sandboxes, harness interop).