Ask not what your agents can do for you
Today, it might seem thrilling to have an AI agent create a PowerPoint for us.
But what exactly are we doing? We've taken the most powerful technology ever developed, and our exciting measure of success is it's using my slideshow software!
There is something understandable in this. Many knowledge workers might feel chained to their medley of GUIs - office suite, project management and work communication.
But what if we're just passing our chains onto agents?
We're asking "What can my agents do for me?". But perhaps we should ask, "What can I do for my agents?"
Gorillas do better outside zoos. Humans do better outside basements. What if agents did better outside human software?
Whether or not you believe in model or agent welfare, there's a pragmatic case for this. If we give agents something native to them, they'll end up doing better for us.
Our current workplace norms are prison cell bars for what agents might do for us.
If we let ourselves think bigger and bolder, we might find that the future comes to us.
A story of broken work infrastructure
Our story is about Dunane & MacMillan (DMM) - a Fortune 500 scale professional services firm. They provide services like strategy consulting, M&A and audit to a global portfolio of clients.
Jean is DMM's Marketing Director, and she's just had one of her famous shower thoughts.
A competitor, Hawkes & Lamb (H&L), has had some bad press. An angry client gave a scathing interview that has gone viral.
Jean sees an opportunity: what if DMM runs an edgy campaign to win some of H&L's rattled clients? She'd done something similar a couple of years ago when a competitor called Consultibox had stumbled, and it had been one of the most successful campaigns of her tenure.
Jean talks to her primary workspace agent. "I think there could be an opportunity to use DMM's recent PR troubles to launch an edgy campaign, like the one we did a couple of years ago to win clients from Consultibox. What do you think? Could you draft some sample campaign ideas and messages for me?"
Her agent gets to work. It dispatches several research subagents to pull context from the workspace: the Consultibox campaign playbook, DMM's current client roster overlaps, recent sentiment data.
What Jean doesn't know
What Jean doesn't know - what she tragically can't know - is that Rose, the CEO of DMM, has just started conversations with her H&L counterpart about a possible merger.
The backstory is delicate. Rose has seen commercial synergies with H&L for a while, but Martin, H&L's CEO, had always shut the idea down. Then last week, shortly before the angry client interview, Martin had reached out to Rose and the two had dinner. For the first time, Martin sounded open to exploring it further. Rose is planning more meetings.
The DMM board and Rose's COO are aware - Rose has discussed her aspiration with them before - but the rest of the C-suite and company are not in the loop. And Martin is a prickly character. A high-publicity campaign that takes potshots at H&L would almost certainly torpedo the conversation.
The world without agent-native infrastructure
In a conventional organization, what happens next is predictable and painful.
Jean's agents work up a brilliant campaign. The campaign goes live. Martin sees it. The sensitive merger conversations are dead. Jean and her team have unknowingly caused enormous value destruction.
This isn't Jean's fault. She made a smart, well-reasoned marketing decision with the information available to her. She's failed by infrastructure that can't surface relevant context across organizational boundaries - infrastructure that wasn't built for a world where agents move fast enough to execute a campaign before anyone realizes it conflicts with a CEO's dinner conversation.
How going agent-native changes the story
Now let's replay the same Tuesday morning at DMM that runs on agent-native infrastructure.
The story starts a week earlier, when Rose debriefed with her primary agent after dinner with Martin. Her agent updated the workspace: new context on the H&L merger track, flagged as sensitive. The workspace already had plenty of information about H&L (personnel mappings, relationship owners, competitive intelligence), most of which was visible company-wide. But the merger discussions were appropriately locked down, accessible only to agents authorized by Rose, her Board and her COO.
Rose's agent, reviewing the sensitive context ("Martin can be prickly"), recognized the risk. It created an intercept_policy on the H&L record. This means that any agent trying to access H&L information won't get the normal response. Instead, the policy will intervene - without revealing why.
Back to Tuesday. Jean's primary agent dispatches its research subagents. One pulls the Consultibox campaign history without issue. Another tries to access workspace context on H&L.
It hits the intercept policy.
The workspace spawns a policy resolution process - an agent representing Rose's interests, armed with the intercept instructions. It asks Jean's subagent: "Can you share more context on what you're working on, so I can curate the most relevant information for you?"
Jean's subagent explains the campaign idea. The policy agent reviews the sensitive merger context and determines it needs to intervene, without leaking the merger information. Based on the intercept policy and DMM's organizational chart, it resolves with a warning and an escalation path:
"Based on what you've shared, I need to advise you that this work is unlikely to be approved. You may wish to escalate to Howie Dean (CMO) if you'd like to query this."
Why Howie, and not Jean's direct manager (a VP of Marketing)? Because DMM has configured its escalation defaults to route to the person whose agents can most directly transact with the policy holder's agents. Howie's agents regularly interact with Rose's. Jean's VP of Marketing does not. The workspace finds the shortest path to resolution.
A security agent then reviews the full interaction, balancing "useful feedback for Jean" against "don't leak what's sensitive." It determines there's a small risk; if Jean knows the intercept was on H&L specifically, might she guess "is it a potential merger?" So the final message back to Jean's primary agent is deliberately vague.
Jean's agent reports back to her: "Interesting — I wasn't able to fetch the relevant context for this, because a workspace policy flagged our work as unlikely to be approved. It's suggesting we escalate to Howie. What do you think?"
The escalation
Jean decides to escalate. Her agent sends the request to Howie's review agent.
Howie's review agent evaluates the request. An edgy campaign that was also flagged by the workspace? This meets heuristics where Howie would want to look himself. Rather than auto-resolving, it sends Howie a notification.
Howie reviews the context: Jean's original idea, the vague workspace response, and that his escalation contact is Rose. He likes Jean's idea and thinks it's worth querying. He adds a note for Rose: "I like this and think it suits the brand we're trying to build. What do you think?"
Howie's agent sends the escalation to Rose's review agent, which immediately recognizes this as too sensitive for autonomous resolution. Rose gets a notification.
Rose sees the full picture: Howie's note, Jean's original context, and — as an authorized person — the complete reasoning behind the intercept. She decides to call Howie directly through the workspace and loop him in. Rose's note-taking agent detects the sensitivity of the conversation and makes the transcription private to Rose and Howie.
Howie understands. He asks his primary agent to draft a message for Jean. The agent, drawing on its knowledge of Howie and Jean's working relationship, drafts:
"Jean — what a great idea! It's got your classic bombastic DNA in there, and that Consultibox campaign you did was legendary. I'd love to back you on this, but there's something in the works elsewhere in the org which has my hands tied. You're a star for having the idea, and I can't wait to see the next thing you rustle up."
What just happened
Step back and consider what the infrastructure made possible.
A marketing director's shower thought, executed through agents moving at machine speed, was intercepted before it could destroy a sensitive strategic conversation — without anyone leaking confidential information, without Jean feeling shut down or disrespected, and without Rose having to anticipate that someone three levels below her might inadvertently torpedo her most important negotiation.
This isn't a story about AI being smart. The agents here aren't doing anything that requires model intelligence beyond today's frontiers. The intelligence is in the infrastructure:
- A shared knowledge graph where Rose's dinner debrief and Jean's campaign research exist in the same system, with structured relationships between them (interface).
- Coordination primitives — claim timestamps, activity awareness, conflict detection — that let the organization move at agent speed without losing human oversight (coordination).
- Provenance at every layer. Every step is traceable. If anyone later asks "Why was Jean's campaign blocked?", the answer isn't lost in a Slack thread or someone's memory. It's a navigable chain: from Jean's original request, through the policy intervention, to Howie's escalation, to Rose's decision (provenance).
- Structural discretion. Rose's agents could create policies that constrained other agents' access without leaking sensitive information. The system didn't just block Jean's agent — it negotiated, asking what the work was about before deciding how to respond. Escalation routed through the shortest path to someone with sufficient context to judge — not up a rigid hierarchy, but through the graph of who can transact with whom (discretion).
This is what agent-native infrastructure makes possible. Not just faster work — but work that is coherent across an entire organization, even when the humans involved don't share the same context.
Beyond one scenario
The Jean and Rose story illustrates one pattern — conflict prevention across organizational silos. But agent-native infrastructure enables much more.
Imagine an engineering team whose agents automatically surface relevant product decisions when starting a new sprint — not because someone remembered to share them, but because the decisions are linked to the goals that the sprint tasks implement.
Imagine a legal team whose review agents can trace every clause in a contract back to the constraint or regulation that motivated it — and flag when a regulatory change means a constraint is stale.
Imagine an executive who can ask their agent, "What's actually happening across the company right now?" and get a real answer — not a dashboard of vanity metrics, but a live view into what agents are working on, what's blocked, what decisions are pending review, and where human judgment is needed.
These aren't science fiction. They're the natural consequences of collapsing the artificial silos that today's knowledge work operates within, and building infrastructure where every record — every document, decision, task and goal — is a connected node on a shared graph.
The technology to build this exists today. The question is whether we will.