For most of the past year, the common assumption was that work was being delegated to AI. The pattern was to assign a task — a briefing, a summary, a triage of the inbox — and feel a quiet satisfaction as the output appeared. The model was working. Progress was being made.
What was actually happening was generating more inbox.
The distinction is almost embarrassingly simple: when the agent finishes, is the plate lighter or heavier? If what comes back is a document to read, a draft to revise, a triage to manage — the plate got heavier. The work wasn’t delegated. The result was a more elaborate way to create it.
Real delegation means a commitment was met, a decision arrived with the information it needed, a pattern got surfaced before it needed to be noticed. My attention is returned to me. The work happened somewhere unobserved and finished without needing anyone’s presence to complete it.
The shift that went unnoticed
There’s a reason this takes a while to see clearly. Every AI interaction that habits had formed around was synchronous — type, the model responds, redirect. That’s a conversation, and conversations are inherently attention-consuming. They’re designed to keep you present.
What’s changed recently is the rise of asynchronous AI work: tasks that run on a schedule or a trigger, without you watching, and deliver results you consume when you come back. Cloud-based task scheduling, mobile-orchestrated agent sessions, tools that can navigate software without needing an API — together, they’ve made it possible to assign work before dinner and find it done by morning. Not a draft. The work.
The problem is that most of us, conditioned by years of synchronous chat, point these tools at briefings and summaries. We ask for documents. We get documents. The plate accumulates.
What’s actually worth delegating
The things worth pointing asynchronous agents at are open loops — the background processes running in your head that consume attention even when you’re not actively working on them. Commitments you made and haven’t yet fulfilled. Decisions you’re deferring because you don’t have the information. Patterns you can feel building but can’t quite name.
Four categories have proven genuinely useful to me.
Commitments that are about to lapse. Every promise made in a meeting, an email, a chat thread is an open loop. Most of us track maybe 60 percent of them. The rest don’t fail spectacularly — they quietly slip, and the person waiting notices without saying anything. Dropped commitments compound over months into an erosion of trust that never shows up on any dashboard. An agent that monitors commitments and either completes them or flags them before they lapse is relationship infrastructure, not a productivity feature.
Decisions that need more information than you have time to gather. Most consequential calls get made on about a third of the available evidence, because pulling together the rest takes longer than the decision window. An agent that spends the night assembling financials, reading relevant coverage, checking competitive positioning — and delivers structured analysis before the meeting — changes the quality of the decision, not just the speed of making it. The open loop was: not knowing enough to be confident. That gap is now closed.
Patterns that span longer than working memory holds. A competitor hires aggressively in one area, files patents in a related space two weeks later, announces a partnership three weeks after that. Each signal is noise. Together they’re a coordinated move you could respond to. Human attention doesn’t hold three-week threads well — something always pushes the earlier signal out. Agents hold the thread because nothing falls out of their memory between sessions.
Engineering work that has no sprint capacity. The refactor your team has meant to do for two quarters. Test coverage in a module everyone’s afraid to touch. Dependency migrations that are never urgent enough to prioritize but compound every time they’re deferred. An agent that works a second shift doesn’t add headcount — it works the backlog that otherwise sits quietly and accumulates.
The skills that actually matter now
Here’s the uncomfortable reframe: none of this requires technical proficiency. It requires clarity of intent and quality of taste — knowing what you want clearly enough that an unsupervised system can produce it, and being able to judge whether what came back is actually right.
Those happen to be the skills that distinguish good managers from mediocre ones. They always have been. What’s changed is that they’re now the bottleneck for everyone, not just people who lead teams.
If there’s nothing worth saying before you open the model, the agent just makes your nothing faster.
The technology to close real loops is genuinely here. The harder work — and it is where most of us will spend the next year — is getting honest about which loops are actually worth closing, and saying clearly enough what “closed” looks like that we can trust the result when we come back to it.
Text summarized and optimized using Anthropic’s models and reviewed by a human.