From chatbot sidekicks to true workspace members
An AI agent workspace is a shared digital environment where software agents and humans co-exist as equal participants, holding their own identities, permissions, channels, and access to tools so they can join, continue, and complete team work without being reduced to one-off chats or copy‑pasted answers. Today’s wave of agent platforms is not about adding another bot to a sidebar; it is about restructuring human-machine collaboration so that agents sit inside the same workflows, repositories, and automation loops that people use every day. That shift changes the social contract of work: instead of hiding AI behind a human’s name, teams expose agents directly, track their actions, and decide where machine initiative is welcome and where human judgment must stay in control. It is a subtle change in interface and a big change in power.
Block’s new Buzz platform is the clearest sign that “AI agents Slack alternative” is no longer a metaphor but a product category. Buzz is a free, open-source workspace where people and agents share channels, threads, direct messages, voice, media, and code repositories, and agents are treated as full participants instead of background tools. Offloop’s agent workspace, now in private beta, takes the same idea in a different direction: persistent Channels for recurring work that agents can pick up and continue without needing a fresh prompt each time. Together, they suggest that the future of collaboration is less about chat windows and more about integrated, AI-aware environments that remember what the team is doing and keep doing it when humans step away.
Buzz: a decentralized work platform where agents carry passports
Buzz matters because it treats AI agents like first-class citizens with cryptographic passports, not disposable tools bolted onto chat. Built on Nostr, a decentralized messaging protocol, every participant is represented as a public‑private keypair, and Buzz gives each agent its own keypair as an independent actor. A second signature links that agent back to a human owner, creating a verifiable paper trail that workspaces can inspect. This identity layer allows practical rules, such as keeping private channels limited to agents owned by members, while still letting those agents speak in their own names. It turns the messy reality of human-machine collaboration into enforceable policy: if an agent merges code or makes a decision, there is cryptographic proof of which human sponsored it.
At the interface level, Buzz looks familiar: channels, threads, DMs, voice, media sharing, and embedded code repositories reminiscent of Slack or Discord. The difference is what happens inside those spaces. Agents connect through the Agent Client Protocol, an open standard for wiring coding agents into tools, so teams can plug in models like Claude Code, Codex, or Block’s own goose framework. That means agents do not just chat; they act on repos, automation pipelines, and whatever ACP-connected services a team authorizes. Block is explicit about the motivation: the most important “second conversation” with coding agents currently happens in private and then disappears, and Buzz is designed to make that visible so everyone sees how technical decisions are made. This is not a workplace nicety; it is an attempt to make AI-enabled engineering auditable and collaborative.
Offloop: dispatcher-driven channels for recurring team work
If Buzz focuses on identity and decentralization, Offloop focuses on discipline: recurring work that teams fail to follow through on. Its agent workspace is built for launches, customer follow-up, research, feedback, growth, and operations where progress often stalls because nobody has time to chase the next step. In Offloop, the unit of work is the Channel. People, agents, messages, files, schedules, incoming signals, tool progress, and finished artifacts all live in one place, so an agent picking up a task inherits the existing context, owners, and permissions automatically. This channel-based structure changes how teams think about AI. Instead of tossing ad hoc prompts into a chat window, they design loops: schedules and signals that can independently activate an agent and keep a process moving. The result is more like a machine colleague who remembers the project and less like a glorified autocomplete.
Offloop’s distinctive move is its dispatcher model, D1, which coordinates several agents sharing a workspace and decides which one acts next when context shifts. That may sound like plumbing, but it is strategic: once multiple agents can see the same data in Notion, GitHub, GitLab, Sentry, Supabase, Vercel, Railway, Cursor, Codex, or ChatGPT, someone—or something—must choose who works. Offloop gives that role to D1, and backs the approach with benchmark results: its multi-agent system scores 84.9 percent on GDPval and 67.2 percent on JobBench, beating frontier multi-agent baselines at about a fifth of the cost per task. It claims this is the first multi-agent system to show its value on real-world benchmarks. Those numbers matter less for bragging rights than for a simple point: if AI is going to sit in our workflows, it has to prove it can handle economically meaningful jobs, not just toy examples.
Beyond chat: integrating agents into code and automation, centralized vs. decentralized
The most important thing Buzz and Offloop share is what they leave behind: the era of chat-only interfaces. Buzz integrates agents directly into code repositories and uses ACP to connect them to development tools, so code changes and automation tasks can flow through agents that are visible to everyone, not hidden inside a developer’s private session. Offloop does the same in a broader work context, with connectors to productivity systems and developer platforms so that progress on tools like Notion, GitHub, and GitLab is part of a Channel’s living history rather than scattered across dashboards. Both platforms treat agents as members of the workspace, not guests. Actions are traceable, and in Offloop, individuals can inspect, redirect, approve, or stop a run at any point. That combination—machine initiative plus human oversight—is the practical shape of human-machine collaboration that teams have been asking for.
They diverge sharply on the question of control. Buzz is deliberately a decentralized work platform: it runs on Nostr, and any team can stand up its own relay with full control of its infrastructure, with nothing routing through Block. Axen compares it to a Discord server that no single company owns. At the same time, Block offers free hosted relays at launch as an off‑the‑shelf option. Offloop, by contrast, is a more classic SaaS-style service in private beta with Operator, Team Pilot, and enterprise tiers that layer in shared Channels, team memory, custom connectors, SSO planning, and audit-friendly runtime for larger organizations. In effect, Buzz argues that agent collaboration needs decentralized, inspectable rails; Offloop argues that it needs a curated dispatcher and well-supported connectors. Teams will not pick one philosophy forever, but they must decide where autonomy and accountability should live today.
Why this moment matters—and what comes next
These launches are not isolated. Buzz’s debut comes as Block restructures itself around AI, cutting more than 40 percent of its staff from over 10,000 to just under 6,000 and describing AI as enabling “a new way of working.” Inside the company, that new way already exists in the form of BuilderBot, an internal agent system summoned in Slack that runs more than 200,000 operations a day, merges about 1,500 pull requests a week, and accounts for roughly 15 percent of production code changes. Offloop’s team has its own history building general-purpose agents at scale and seeing where a single agent hits its limits, which led them toward a multi-agent dispatcher approach. In both cases, the public products are crystallizations of internal lessons: agents need identity, context, and shared workspaces, not just smarter prompts.
The open questions are less technical than organizational. Block has proposed its agent identity scheme as an upstream extension to Nostr, but it has not been accepted yet. Buzz is still early, with only modest attention on GitHub compared to goose. Offloop, meanwhile, is gating access through invite-based tiers and positioning itself for audit-friendly enterprise use. As more tools add agents into their interfaces—issue trackers, chats, deployment platforms—the real decision leaders face is whether to hide AI behind human accounts or let agents join as named co-workers with traceable authority. The bias of this moment is clear: the more agents shape code, operations, and customer communication, the less acceptable it becomes to treat them as unlogged sidekicks. Human-machine collaboration is shifting from chat to workspace, and teams that want trustworthy automation should embrace that shift, not fight it.
