Anthropic's second post in its human-agent teams series is an interview with Slack's Chief Product Officer, Jaime DeLanghe, on how Slack thinks about embedding agents into the flow of workplace conversation. It's a discipline the company was already practising with humans long before AI arrived.
DeLanghe joined Slack in 2017 to work on search and machine learning, with a mission to turn workplace conversation into institutional knowledge. Her core belief is that getting there requires people to work in the open, keeping conversations, decisions, and work-in-progress in channels anyone at the company can read and search, rather than locked in direct messages or someone's head.
Her essay "The Work is the Conversation" makes the same case for agents. The conversation happening around a piece of work is exactly the context an agent needs to be useful. Without it, an agent is operating blind. It can execute a task, but it can't tell you why the task exists, what's already been tried, or who to loop in when something looks off.
Why conversation didn't become knowledge, until now
For years, the promise that workplace conversation would compound into organisational knowledge simply didn't materialise. As DeLanghe puts it: "I have so many research papers from the early days at Slack that showed that, actually, no, conversation doesn't turn into knowledge. You wish it did, but really it's just a lot of stuff that goes past."
That's an unusually honest thing for a product leader to say about their own category. The reason it didn't work is straightforward: a human skimming a busy channel can't hold a year of context, and search only helps if you already know roughly what you're looking for.
Agents change the equation. Unlike a person catching up on a channel, an agent can read, search, and act on that conversational exhaust in real time. It can surface a decision made three weeks ago in a different channel, or flag that someone already tried the fix you're about to attempt.
What this means for Australian businesses running Claude
The practical lesson isn't about Slack specifically. It's about where you let your agents see.
If Claude only has access to a narrow slice of context, a single ticket, a single document, a single inbox, it will keep re-deriving decisions your team already made. You get a competent assistant with amnesia. The businesses getting the most out of Claude Cowork and Claude Code are the ones treating existing conversation and documentation as an asset the agent should be able to search, not a private archive to be protected from it.
For Australian SMBs this has a concrete shape: connecting Claude to the tools where decisions actually happen, rather than restricting it to a single upload-and-ask workflow. In practice that means:
Team chat, where most real decisions get made and then immediately forgotten.
Email threads with clients and suppliers, which usually hold the commitments nobody wrote down anywhere else.
Project trackers and CRM notes, where the current state of a piece of work actually lives.
Your document store, so the agent is reading the live version of a policy rather than a copy someone pasted in six months ago.
Working in the open is a prerequisite, not a nice-to-have
There's an uncomfortable implication in DeLanghe's argument. If your organisation runs on direct messages and private threads, an agent connected to your workspace will see very little of substance. The context that makes it useful is precisely the context your team has been keeping out of shared channels.
That's a cultural problem before it's a technical one, and it isn't solved by buying anything. But it's worth naming early, because it explains why two businesses can deploy the same tool with the same connectors and get wildly different results. The one whose team already works in the open gets an agent with institutional memory. The other gets a chatbot.
Where the access boundary should sit
Broad read access is the point, but it isn't unconditional. Anything covered by the Privacy Act, HR matters, personal information about staff or clients, commercially sensitive negotiation threads, deserves a deliberate decision rather than a default.
The workable pattern for most Australian businesses is to open up operational context generously and fence off the genuinely sensitive channels explicitly. That's a short scoping conversation, not a project, and doing it upfront avoids the much worse alternative of discovering the boundary after something has already surfaced in the wrong place.
From a single pilot to a connected rollout
The upside DeLanghe describes, scattered knowledge finally compounding into productivity, is the same upside worth scoping when a business moves from a single Claude Code pilot into a properly connected Cowork rollout. The difference between those two states is almost entirely about connectors and permissions, not about the model.
A connected rollout of this kind, covering connector setup across your main systems, an access policy, and a short training pass so the team knows what the agent can now see, typically runs A$3,500 to A$7,000 for a small Sydney or Melbourne business depending on how many systems are in scope.
The piece is the second in Anthropic's series on human-agent teams. The first covered lessons from building teams with multiplayer AI inside Anthropic itself, and both are worth reading if you're weighing how far to open up your own context.
If Claude is currently working from whatever someone remembers to paste into it, that's the gap worth closing first. Book a session and we'll map which of your systems it should be reading.
