AI agents need a workroom, not another chatbot

The most interesting AI product I saw this week was not a model.

It was a room.

Block has released Buzz, a free open-source workspace where humans and AI agents sit in the same channels, share context, discuss work, review code and run approved workflows. It is built on Nostr, so humans and agents get cryptographic identities rather than just being anonymous bot accounts inside somebody else’s SaaS box.

That sounds a bit nerdy. It is. But the underlying point is simple.

Chat was fine when AI answered questions.

Chat is not enough once AI starts doing work.

If an agent can read your files, inspect a codebase, update a CRM, draft a proposal, publish a WordPress post, open a pull request or touch customer data, the interface cannot just be a private conversation between one human and one model. That is too small. Too hidden. Too easy to lose.

You need to know where the work happened. You need to know which agent did it. You need to know what context it saw, what tools it used, what it changed, what failed, what was approved and who owns the decision.

That is the part most AI demos still skip.

They show the clever bit: “look, it shipped a thing.” Fine. We have all seen the magic trick by now. The more useful question is what happens after the trick.

Can another person review the work without digging through a private chat? Can the business see the evidence? Can you tell the difference between a draft, a suggested action and a live change? Can you revoke the agent’s access without praying it has not inherited half your local environment? Can you replay the run when something goes wrong?

That last point is not theoretical. The Buzz demo I looked at was impressive because it was low friction: install the app, connect existing Claude Code and Codex subscriptions, create channel-scoped agents and ask them to work. One agent inspected local files and eventually published a WordPress post through credentials already sitting in the operator’s environment.

That is both the appeal and the red flag.

Nobody wants another six-week enterprise setup before an agent can draft a blog post. But if a workspace inherits local files, environment variables, WordPress access and existing workflows too freely, the productivity story turns into a permissions story very quickly.

This is where the market is heading.

OpenAI’s Presence product is aimed at enterprise agents that resolve customer and internal workflows with policies, approved actions, evaluations, escalation rules and post-launch improvement. OpenAI’s Codex research says agent work is already moving from short chat interactions into long-running delegated tasks, including non-developer work across legal, finance, recruiting, marketing and operations.

Anthropic is pushing the same direction from the model side. Opus 5 is pitched around steadier long-running coding, automation, computer-use and professional work. Not a toy prompt. Work.

Then look at the operator end of the market. Orca is trying to become the visual control plane for parallel coding agents: worktrees, terminals, browser sessions, diff review, mobile supervision. Ryan Carson’s AI software factory is not really a story about replacing engineers with magic. It is a story about queues, evidence, model routing, production watchdogs, browser QA, cost control and a human still making the calls every 25 minutes.

Different products. Same direction.

The agent is not the whole system. The room around the agent is becoming the product.

For business owners, this matters because the lazy version of AI adoption is now obvious: buy a few subscriptions, tell people to use them, maybe bolt a chatbot onto the website, then call it transformation.

That is not transformation. That is tool sprawl with better autocomplete.

A real AI operating layer needs a place where work can be seen and governed. Not a theatrical dashboard nobody opens. Not a private prompt history. A proper working surface.

At minimum, that surface needs eight things.

First, named agents. Not “the AI.” A research agent, a support agent, a draft agent, a QA agent, a pricing agent. Each with a job.

Second, scoped permissions. The research agent probably does not need publish access. The draft agent does not need billing access. The support agent should not quietly discover it can change product pricing because somebody reused an admin token.

Third, shared context. The agent should see the relevant brief, files, customer history, policy and current task. It should not be rummaging around the whole business like a bored intern with root access.

Fourth, visible history. If the agent made a recommendation, changed a file or triggered an action, the business needs the run history. Timestamps, inputs, outputs, tool calls, approvals, errors.

Fifth, evidence. Screenshots, links, diffs, test results, source references, customer records. “Done” is not evidence. “Here is what changed, here is how I checked it, here is what still worries me” is evidence.

Sixth, approval gates. Some tasks can run automatically. Some should stop at a draft. Some should require a named person to approve the next step. The system should enforce that, not politely suggest it.

Seventh, rollback. Agents will make mistakes. Humans do too. The difference is that agents can make mistakes at speed, at night, across several systems, while sounding horribly confident. If you cannot pause, revert or contain the work, you are not ready for much autonomy.

Eighth, an owner. Every agent needs a human owner. Not because humans are always smarter. Because the business needs accountability. If the agent sends the wrong quote, publishes the wrong claim or changes the wrong setting, somebody has to own the recovery.

This is where the agency model changes.

The old agency sold output: campaigns, websites, content, reports, ads, decks. The slightly updated agency sells AI-assisted output, which is mostly the same thing with faster drafts and worse taste if nobody is paying attention.

The post-agency company sells the operating layer.

It diagnoses the commercial workflow. It builds the research loop, the campaign loop, the content loop, the sales follow-up loop, the customer support loop. It wires the agents into the right parts of the business, with the right permissions, evidence and approvals. It does not pretend the model is the product.

The product is the managed system around the model.

That is the practical opportunity for smaller businesses too. They probably do not need OpenAI Presence. They do not need to self-host Buzz tomorrow. They definitely do not need to hand a random agent access to everything because a demo looked tidy.

They need a much simpler first pass: pick one workflow, map what happens today, decide what the agent may read, draft, recommend or change, run it in shadow mode, compare the output against real work, track errors, add approvals, then widen the scope only when the evidence says it is safe.

This is less exciting than a viral agent demo.

Good.

Most businesses do not fail with AI because the model is not clever enough. They fail because nobody designed the room where the work was supposed to happen.

---

Pull quotes

• Chat was fine when AI answered questions. It is not enough when AI starts doing work.
• The agent is not the whole system. The room around the agent is becoming the product.
• “Done” is not evidence. Show the source, the diff, the test, the approval and the thing that still worries you.
• Tool sprawl with better autocomplete is not transformation.
• The post-agency company sells the operating layer around the model.

Short LinkedIn / X version

The most interesting AI product this week was not a model. It was a room.

Block’s Buzz puts humans and AI agents into the same workspace: channels, threads, code, workflows, agent identities and permissions.

That is the right direction.

Chat was fine when AI answered questions. It is a poor interface once AI can read files, touch CRM data, publish content, open PRs or update customer records.

The business question is shifting:

Not “which model?” Not “which chatbot?” But “where does the work happen, who can see it, and who approved the action?”

A serious agent setup needs named agents, scoped permissions, shared context, visible history, evidence, approval gates, rollback and a human owner.

Otherwise it is just tool sprawl with better autocomplete.

This is the post-agency shift: the product is not AI-generated output. The product is the operating layer around the work.

Notes and caveats

• Buzz is labelled a developer preview. Treat it as a test surface, not production infrastructure.
• The Buzz account here is based on a locally reviewed demo plus Block’s launch materials and repository. The WordPress publishing behaviour was observed in that demo, not independently reproduced in a sandbox for this piece.
• Block’s launch post says agents have cryptographic identities and defined permissions. That does not, by itself, prove production-grade least privilege, secret isolation, audit retention or revocation behaviour.
• OpenAI Presence is limited general availability for eligible enterprise customers and is not self-serve.
• OpenAI’s Codex work-duration claims are model-estimated and should be treated as directional, not exact labour accounting.
• Anthropic Opus 5 claims are vendor claims and benchmark/customer-report based. Useful signal, not neutral proof.
• CTA option if this runs with a direct offer: “If you are adding AI to marketing, start by mapping the room: context, permissions, evidence, approval and rollback. The model comes after that.”