Murph's Take

Why Background Agents Change Everything About How You Structure Work

When AI runs in the background — not waiting for your prompt, but working on defined tasks while you do something else — the structural implications for how work is organized are profound and underappreciated.

Jason MurphyMarch 29, 20267 min read

There's a friction that's so embedded in how we think about knowledge work that most people don't notice it: the work we do requires our continuous presence. Thinking happens when we're thinking. Writing happens when we're writing. Research happens when we're researching. The work is synchronous with the worker.

Background agents break this assumption. And the implications are more structural than most people's current framing captures.

The Synchronous Assumption

When you use Claude in a standard chat interface, the assumption is synchronous: you're present, you send a message, you wait for the response, you decide what to do with it, you send the next message. The interaction unfolds at human pace. The AI is waiting on you at least as much as you're waiting on it.

This is better than doing the work entirely alone. The outputs are higher quality, the process is faster, you can cover more ground in a given session. But the fundamental architecture hasn't changed: you're still the process manager. Work proceeds when you're there to direct it and pauses when you're not.

Background agents — Claude Code in background mode, properly scoped agentic processes, autonomous loops — remove that assumption. The work continues when you're not there. You return to a finished state rather than a paused one.

What This Actually Enables

The naive framing is "AI does work while you sleep." That's true but undersells the structural change. The more accurate frame is: you can now work on multiple things simultaneously because execution is no longer coupled to your presence.

In a traditional knowledge work structure, multitasking is largely an illusion. You context-switch between things, which means each thing moves slowly and the quality of each is degraded by the switching cost. Gloria Mark's research at UC Irvine estimates 23 minutes of recovery time per interruption — genuine multitasking is cognitively expensive.

Background agents enable a different kind of parallelism that doesn't pay that cost. While you're in a focused session on one problem, an agent is executing on a parallel problem. You're not splitting your attention. You're delegating execution so your attention can stay whole.

The structural implication: the number of things you can meaningfully advance in a given time period is no longer limited by your attention bandwidth. It's limited by your ability to design clear tasks, scope them correctly, and review the outputs. Those are different bottlenecks, and they're bottlenecks that compound better.

Design First, Execute Second

Background agents surface a design discipline that's easy to skip in synchronous work. When you're present in a conversation, you can course-correct in real time — if Claude heads in the wrong direction, you redirect. The feedback loop is tight enough that vague instructions often work out.

Background agents don't have that feedback loop. If the task is scoped ambiguously, the agent will complete something. It might be the wrong something. And you won't know until you come back to review.

This makes task design a first-class skill. Before setting a background agent on anything consequential, you need:

A clear statement of the goal and why. What success looks like in specific, observable terms. What's explicitly out of scope. What should trigger an escalation rather than autonomous continuation. What format the output should take.

This is, incidentally, the same discipline that makes human delegation work. The organizations with good management practices are the same ones that will transition smoothly to background agent work, because the underlying skill is identical: clear task design with defined success criteria and explicit escalation conditions.

The Verification Architecture

Background execution without verification is risk without corresponding upside. You need to come back and review what happened. The question is what that review looks like.

Good background agent verification isn't "read everything the agent produced." It's "review the decision log and spot-check the output." I want to see what judgments the agent made, where it had to choose between options, and what it did when it encountered something unexpected. If those are documented clearly, I can review in ten minutes what would take an hour to re-generate.

The output itself I sample. I'm not verifying that Claude wrote good code by reading every line — I'm verifying that the tests pass, the interfaces are consistent with the rest of the codebase, and the structure matches the patterns I established. If those conditions hold, I trust the detail.

The organizations that build good verification practices now will have an enormous advantage as background agent capability increases. The ones that either never verify or verify everything will hit different ceilings.

What tasks in your current work are actually well-suited to background execution — and what's stopping you from designing them that way?

Want this for your business?

Tell us what you're building. We'll map out exactly what to build and what it costs.

Start Your Project →

Frequently Asked

What is a background agent in practical terms?

A background agent is an AI process running autonomously on a defined task while you're doing something else — not waiting for your next message, but proceeding through a multi-step operation, observing results, and making decisions within its scope. Claude Code's background mode is the clearest current implementation: you can kick off a complex task and return to a finished result rather than supervising each step.

How does this change work structure?

Synchronous AI collaboration requires your attention for each interaction. Background agents allow genuine parallelism — you design a task, set it running, and use that time for other work. The structural shift is from 'AI makes my work faster' to 'AI allows me to work on more things simultaneously by handling execution while I focus on design and judgment.' That's a different relationship to time and capacity.

What kinds of work are well-suited to background execution?

Work with clear success criteria, defined scope, and minimal need for real-time judgment calls. Code refactoring within defined patterns, research compilation with specified sources and format, content drafting to a detailed brief, test suite generation, documentation from code — these translate well to background execution. Work that requires ongoing judgment about subjective criteria, novel decisions with high stakes, or real-time feedback loops is better done synchronously.

What's the failure mode to design against?

Background agents that operate without sufficient constraint produce outputs that are locally correct and globally incoherent. The failure is usually insufficient scope definition — you wanted the agent to refactor the authentication module, it refactored the authentication module and incidentally changed three things that affected the payment flow. Good background agent design includes explicit boundaries, verification criteria, and escalation conditions that trigger a return to human judgment when something is outside the expected scope.

Jason Murphy

Written by

Murph

Jason Matthew Murphy. Twenty years building digital systems for businesses. Former CardinalCommerce (acquired by Visa). Now running VibeTokens — AI-built websites and content for small businesses.

The window is open.

It won't be forever.

Start Your Project →