Async AI Agents Are Your New Marketing Intern

The way founders use AI is splitting in two. One group still types into a chat window, waits for a reply, copies the output, and pastes it into the next tool. The other group writes a goal, walks away, and comes back hours later to finished work sitting in a pull request.

The split is the whole game now. The founder-marketers who learn to brief async agents like a junior hire are going to outpace the ones still treating AI as a synchronous chat tool, and the skill that separates them is goal-writing, not prompt-engineering.

Engineering normalized this first

Lenny Rachitsky covered the new /goal feature in OpenAI Codex this spring, the one engineers are calling “the Codex feature that works while you sleep.” The framing is straightforward: write a goal spec, the agent runs in the cloud for as long as it needs, you wake up to a draft pull request. The thesis on Lenny’s Newsletter lays out a six-part structure for the goal itself, but the bigger move is the shift in interaction model. You stop holding the agent’s hand turn by turn. You write a tight enough brief that it can run unattended.

That move is not unique to Codex. Anthropic’s Claude Code documentation now opens with the same posture. The product is described as “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools,” available in the terminal, the IDE, the desktop app, and the browser. The web version sells itself on exactly the same async muscle: “Kick off long-running tasks and check back when they’re done, work on repos you don’t have locally, or run multiple tasks in parallel.” Different vendor, same shape.

A new feature called Routines makes it even more explicit. Anthropic’s docs describe a routine as “a saved Claude Code configuration: a prompt, one or more repositories, and a set of connectors, packaged once and run automatically.” Triggers can be schedule, API call, or GitHub event. The runs happen on cloud infrastructure, “so they keep working when your laptop is closed.” The key line in the docs is the one most people skip: “The prompt is the most important part: the routine runs autonomously, so the prompt must be self-contained and explicit about what to do and what success looks like.”

That sentence is the whole skill, transplanted into marketing.

What this looks like inside a marketing function

Most founder-marketers are still using AI the way they used Google in 2009. Open a tab. Ask a question. Read the answer. Close the tab. Even the ones running Claude Code from the terminal are mostly running it interactively, watching the cursor blink, approving each step.

The async shift changes the unit of work. The unit stops being a prompt and starts being a goal with a definition of done. A few examples of what that looks like in a 5-30 person B2B SaaS marketing motion.

• Weekly competitive sweep. A scheduled routine pulls the last seven days of changelogs, blog posts, and pricing-page diffs from your top six competitors, then writes a Monday brief in your CMS draft folder. You read it with coffee.
• Newsletter draft from raw inputs. A goal points at a folder of customer-call transcripts, three product release notes, and the last newsletter for tone. The agent ships a draft into a Google Doc named for next Thursday’s send.
• Inbound lead enrichment. A trigger fires when a new lead lands in your CRM. The agent looks them up, summarizes the company in two sentences, drafts a personalized first reply, and attaches both to the record.
• SEO topic backlog rebuild. Once a month, the agent pulls Search Console data, cross-references your existing posts, flags topical gaps, and opens a pull request against topic-backlog.md with ten ranked suggestions and citations.

None of these are theoretical. The Claude Code routines docs cite “alert triage,” “backlog maintenance,” “deploy verification,” and “docs drift” as the canonical use cases. Swap a few nouns and they are marketing workflows. The vendor wrote the patterns for engineers because engineers were the early adopters. The patterns themselves are general.

The skill is goal-writing

The reason most marketers’ first attempts at async agents fall flat is the same reason most first managers’ first hires fall flat. They under-brief.

A good async brief looks more like a junior-hire onboarding doc than a prompt. The goal has to be self-contained, since nobody is in the room to clarify halfway through. It has to declare what success looks like, since the agent has to know when to stop. It has to specify the artifacts: a draft in a folder, a pull request against a branch, a Slack message to a channel. It has to handle the obvious failure modes: what to do if a source URL is dead, what to do if the data is empty, what to escalate vs. what to defer.

This is closer to writing a job description than writing a prompt. Which lines up with the framing in agents are your new employees: the founders winning right now are the ones who built a manager relationship with their agents, not a tool relationship. The async version raises the stakes on that. A synchronous chat lets you correct course mid-conversation. A scheduled routine that ran at 3 a.m. has already committed.

A practical brief template that has been working for me:

1. Role and context. Who the agent is acting as. What the brand voice is. What folder or repo it lives in.
2. Trigger and cadence. When this runs and what kicks it off.
3. Inputs. Exact files, URLs, or data sources it can read.
4. Process. The steps in order, with named tools or MCPs where relevant.
5. Output spec. What the artifact is, where it goes, what filename, what format.
6. Definition of done. The checks that have to pass before the agent considers itself finished, and what to flag for a human if they don’t.

That last one is the part marketers tend to skip. The agent needs a rubric, not just a task. “Three sources cited” is a rubric. “Make it good” is not.

What changes about the org chart

The interesting consequence of all this is what happens to headcount math when a chunk of the work runs while you sleep.

A 15-person SaaS company traditionally has one marketer, maybe a contractor for content, maybe an agency on retainer. Output is bounded by how many hours that one marketer has. The async shift breaks that ceiling. The marketer’s hours now go into the brief, not the deliverable. Every brief that gets written once produces output every week, forever, with the marketer reviewing instead of producing.

The role that lands is not “marketer who uses AI.” It is the operator role I covered in the first 30 days of a Marketing Engineer, with the async layer bolted on. One human writing and reviewing briefs. A bench of agents running the recurring work. A weekly cadence where the human grades the agent output and tunes the brief, the same way a manager grades a junior’s first month.

The companies that figure this out in 2026 will look unrecognizable in 2027. Not because they hired faster. Because the things their org actually does each week stopped being capped by human-hours.

What to do this week

If you have never run an async agent against a marketing workflow, three concrete moves.

1. Pick one recurring task you do every week. Pick the most boring one. The Monday metrics pull, the weekend social calendar prep, the newsletter draft from a template.
2. Write the brief in the six-part shape above. Save it as a markdown file somewhere durable, not in a chat window. Treat it as a living document you tune.
3. Run it once manually, end to end, while you watch. Note every place the agent guessed wrong, every place the brief was ambiguous. Tighten the brief. Then schedule it.

The first agent build will feel slow. The second one will feel faster. By the fifth, the briefs start looking similar enough that you can template them. That is the moment the output compounds.

If you want help wiring async agents into your marketing motion, the AI Opportunity Sprint is the four-week version of that conversation. It’s how I work with founders who are trying to move from chat-window AI to a real agent bench before the rest of their market catches up.

Jared Castronova is the founder of JAC Growth Marketing, where he builds AI-powered GTM systems for B2B companies.