Article 3/4 in the AI Agents series — What is an AI agent? · How to build an AI agent? (coming soon) · AI agents vs. organizations (coming soon)
You know what an AI agent is, and how to build one. What's left is the question nobody asks in the demos: what does it actually do to a team?
This morning, before my coffee, three agents had already worked for me. The first had combed through a hundred-odd sources to prepare my weekly intelligence digest and handed me a sorted, scored file. The second had pushed a module refactor while I reviewed, stopping only at the few decisions that deserved my opinion. The third had done nothing — it was waiting, and a small box on my desk signaled it with a green light: nothing to approve, I could get on with my morning.
None of these three agents is a technical feat. What changed isn't their intelligence — it's their place. They're no longer tools I open when I need them. They've become an execution layer running in the background, one I work with the way I'd work with a small team.
In the first article, I defined what an AI agent is. In the second, I built some. Here, we look at the consequence: what happens when you put these agents inside a team — your own, or the one you lead. And why it's an organizational question long before it's a tooling one.
What agents actually do inside a team, today
Forget the promises. Here's what's already running, in 2026, in ordinary teams.
Code. This is the most mature use case. Claude Code, Mistral Vibe, Codex, Cursor: a developer describes an intent, the agent writes, tests, fixes, opens a pull request. On my own stack, I let an agent drive refactors that I review rather than write. The craft doesn't disappear — it shifts from "typing the code" to "framing, reviewing, deciding."
Intelligence and research. Aggregating, filtering, summarizing, cross-checking sources: long, repetitive work, perfect for an agent. My weekly digest is produced by a chain of agents that collect, deduplicate and score — I keep control of the editorial angle and the final selection, they absorb the gathering.
Content and marketing. Copy variants, visual derivatives, performance analysis: tasks that used to require whole teams now fit on a single, well-equipped person.
Operations. Ticket triage, meeting summaries, database updates, follow-ups: the administrative "execution layer" can be delegated piece by piece.
What these four families share isn't sophistication. It's that, in each one, the part of the work that consisted of producing collapses, while the part that consists of deciding stays whole.
The pattern: from execution to orchestration
Here's the shift I think is underrated.
For decades, a professional's value was measured by their capacity to do: produce content, code a feature, build a spreadsheet, process a case. Execution was scarce, therefore expensive, therefore a differentiator.
AI agents break that equation. When an agent executes specs, aggregates information, produces deliverables and runs workflows without fatigue or friction, the marginal cost of execution trends toward zero. And when execution costs almost nothing, what's left — what creates value — is only the layer above: deciding what to do, judging what's produced, and connecting it all to a goal.
This doesn't mean execution skills become useless. They become implicit prerequisites rather than differentiators. Just as knowing how to read a spreadsheet no longer justifies a role, knowing how to produce content or code a feature will soon no longer be enough unless there's, behind it, a capacity to orchestrate, judge and arbitrate.
The profile that emerges is someone who understands enough tech to configure the right agents, enough domain knowledge to set the right objectives, and enough strategic sense to know when not to delegate to a machine. It's not a new profile: it's exactly what good chief digital officers, good product leads, good operations managers have done for twenty years — except now their operational arms are agents, not teams of fifty.
What it really implies: the uncomfortable question
You might find this analysis abstract. A story made it very concrete in early 2026.
For nearly ten months, the entire growth marketing team at Anthropic — the company behind Claude — rested on a single person. Austin Lau, hired with no technical background, recounts that when Claude Code launched he "had zero idea what this product is for" and had to look up how to open the Terminal on his Mac. On his own, with Claude Code as his operating arm, he ran the ad campaigns, the performance analysis, social, press relations, design. Ad-creation tasks that took half an hour were done in about thirty seconds (a story reported from an Anthropic case study, picked up in early 2026).
One non-technical person absorbed the workload of an entire marketing team, at one of the most closely watched companies in tech. This isn't a productivity anecdote. It's a demonstration of what happens when execution costs zero: an augmented operator doesn't replace a colleague, they replace a function.
That's the consequence "AI deployment" roadmaps carefully avoid looking in the eye. The real question for an organization isn't "how do we add agents to our tools?" It's: which roles are, deep down, disguised execution layers — and which roles hold because of their capacity for judgment?
The traps no demo shows you
Before fantasizing about a team of agents, three walls I've hit — or watched others hit.
Cost that runs away. An agent that works continuously consumes continuously. In spring 2026, Uber had to cap its use of Claude Code and Cursor after burning through its AI budget in four months. The price per token is falling, but the bill is rising, because consumption explodes. Deploying agents without caps, quotas and monitoring is replaying the worst of uncontrolled cloud costs. Financial governance (FinOps) has to arrive before deployment, not after the invoice.
Permission that fades. An agent acting on its own can write, delete, push, send. The question "who approves what, and when?" becomes central. That's exactly the problem I tackled with my physical pager wired into Mistral Vibe: decoupling the developer's presence from the agent's execution, without losing control over sensitive actions. At an organization's scale, this need takes the form of permission layers, human oversight (human-in-the-loop) and, soon, agent registries: the Linux Foundation is already pushing an open, DNS-based standard to inventory and discover an enterprise's agents. You can't govern what you don't inventory.
Judgment delegated out of laziness. The subtlest trap isn't technical. It's letting the agent decide because it's faster, where a decision genuinely commits you. An agent approving its own commits, a support assistant granting access because it's politely asked to (it happened, on Instagram accounts, by simply asking an AI bot) — these aren't bugs, they're badly drawn boundaries. The rule fits in one sentence: delegate execution, never the arbitration that carries consequences.
The real question for your organization
If you lead a team, the right entry point isn't "which agentic tool should we adopt." It's a mapping exercise.
Take your processes, and for each one ask a single question: does this task require irreplaceable human judgment, or is it execution ready to be orchestrated? You'll get two columns. The first — decisions that commit you, ethical trade-offs, relationships of trust, strategic choices — stays human, and that's where you should concentrate your best people. The second — production, aggregation, formatting, follow-up — is a candidate for orchestration by agents, under supervision.
This sorting isn't a headcount cut in disguise. It's a re-reading of what truly creates value in your organization. Most teams will discover they spend a considerable amount of time in the second column — and far too little in the first.
Conclusion: redrawing the do / decide line
Putting agents on a team isn't plugging in one more tool. It's moving the line between what we do and what we decide — and accepting that a professional's value, tomorrow, will be measured less and less by their capacity to execute, more and more by their capacity to orchestrate and judge.
The organizations that nail this transition won't be the ones that deployed the most agents. They'll be the ones that were clear-eyed about the uncomfortable question: in what we do every day, what truly deserved a human?
The next and final article in the series steps up a level: from agents on the team to the governance of agents facing organizations — autonomy, accountability and the strategic questions this shift imposes on leaders.
Spotted a factual error, or want to share your own deployment feedback? Write to me at bonjour@romaindelfosse.fr.
Sources
- picked up in early 2026 — growthtalent.org
