Claude just shipped managed agents — here are the ten I deployed in one afternoon.

It is 9:14 AM on a Tuesday and I have, at this moment, zero marketing agents running.

By 4:30 PM I will have ten. They will run tonight. Some of them will argue with each other — mediated by a Critic agent that does not particularly care about anyone's feelings — and by morning there will be outputs waiting for me. No servers configured. No infrastructure provisioned. No DevOps call scheduled for the following Thursday.

This is what managed agents actually look like in practice: not a whiteboard architecture and not a six-month deployment. An afternoon. And then work.

What follows is an account of how that afternoon went, what each agent does, and — this is the part the marketing glosses over — what each one will categorically refuse to do.

9:14 AM — The Brief

Before any agent is deployed, the brief. This is the non-negotiable precondition that most people skip because they are eager to get to the deployment part. They pay for this later.

The brief is not a prompt. It is a document — maybe eight hundred words — that answers four questions: who are we, who do we serve, what does quality look like in our context, and what are the boundaries of autonomous action? The last question is the one that matters most for managed agents. It defines the escalation threshold: what the agents handle without asking, and what requires a human before anything ships.

With the brief written and the brand policy file updated, the afternoon begins.

10:02 AM — Agents One and Two: The Intelligence Layer

Agent 1: The Morning Analyst. The foundational agent in any marketing stack. Reads overnight performance data — ad accounts, site analytics, conversion events — and produces a structured briefing every morning before 6 AM. Its standing question is not "what happened?" but "what requires a decision today?" The briefing is five sections maximum, ends with a ranked list of recommendations, and is routed through the Critic before it reaches my inbox.

What it will not do: make budget changes autonomously. Every recommendation is a recommendation. The analyst advises; I decide. This is a policy choice, not a technical limitation, and I made it deliberately.

Agent 2: The Competitor Spy. Monitors a defined list of competitor domains on a weekly cycle. Reads public signals — published content, apparent campaign activity, pricing page changes, announcement cadence — and produces a structured weekly report. Categorizes changes by apparent strategic intent: content play, acquisition push, pricing test, product update.

What it will not do: speculate beyond observable signals. I wrote this into its instructions explicitly. The Spy does not offer opinions on competitor motivations. It reports facts and lets me draw conclusions. The temptation to give an intelligence agent an editorial voice is strong; I have learned it produces confident-sounding noise.

11:15 AM — Agents Three and Four: The Content Engine

Agent 3: The Brief Writer. Converts a one-sentence idea or rough outline into a full campaign brief: audience definition, tone guidance, three angles to test, a clear success metric, and a list of questions the brief cannot answer that require a human decision before execution. This last element — the explicit unknowns list — is the feature most Brief Writer implementations skip and the one that saves the most time.

What it will not do: invent audience data it does not have. If the ICP is underdefined for a given brief, it surfaces the gaps rather than filling them with plausible-sounding assumptions. This is the Critic's most common note to the Brief Writer: "you have defined this audience based on assumptions that are not in the policy file." The Brief Writer, appropriately, routes these back for ICP clarification rather than proceeding.

Agent 4: The Copy Engine. Writes ad copy, landing page sections, and email drafts based on approved briefs. Reads the brief. Does not deviate from the audience definition, the approved tone, or the stated angle without flagging the deviation in an explicit note attached to the output. Every output goes to the Critic before it reaches the scheduler.

What it will not do: ship anything the Critic has not reviewed. This is architectural, not instructional. The Copy Engine's output is never a final artifact. It is always an input to the review process.

12:47 PM — Lunch, and an Agent Running Without Me

By noon, four agents are configured and running their first cycles. I eat lunch. The Competitor Spy runs its initial benchmark pass. When I return, there is already a preliminary report in the queue — a note that one competitor has published six pieces of content in the last eight days, all targeting a keyword cluster I am not currently competing on.

This is the part that feels different from using a tool. A tool requires you to ask. An agent tells you what it found.

1:30 PM — Agents Five and Six: Paid Media

Agent 5: The Negative Keyword Hunter. Reads paid search account data weekly. Analyzes search term reports to identify non-converting query patterns that are consuming budget. Groups findings by pattern type. Produces a structured list of negative keyword candidates with match type recommendations and estimated monthly waste per pattern.

What it will not do: apply negative keywords automatically. The Hunter recommends; a human approves; a separate process applies. This is not over-caution — it is the lesson of every automated bid management horror story. Paid account changes without human review are how you accidentally exclude your own branded terms.

Agent 6: The SEO Landing Page Auditor. Audits existing landing pages against target keyword strategy and conversion best practices. Checks title tags, meta descriptions, heading structure, internal linking, page speed signals, and conversion element placement. Produces a prioritized improvement list with severity scores. Runs monthly, with an alert trigger if a page's search ranking drops more than fifteen positions between cycles.

What it will not do: edit pages autonomously. Every recommended change is flagged for human review. The Auditor identifies; humans implement. This boundary is important because landing page changes affect conversion rates in ways that are not always predictable from the audit signal alone.

2:45 PM — Agents Seven and Eight: The Quality and Relationship Layer

Agent 7: The Critic. Reviewing the Critic last is wrong — it is the first thing you should understand about any agentic marketing operation — but explaining it last makes narrative sense because you see what it reviews before you understand why it matters.

The Critic reads every substantive output from every other agent before it routes to a human or to a scheduler. Its brief is fixed: check against brand voice, factual accuracy, legal conservatism, competitive sensitivity, and a gut-check question — would we be embarrassed if this ran? It returns a structured review: pass, revise, or reject. Rejections include specific notes. Revisions go back to the originating agent with the Critic's notes attached.

What it will not do: approve work it has not read. This sounds obvious. It is remarkable how many "agent stacks" route the Critic as an optional step rather than a required gate. Optional gates get skipped when timelines compress. Required gates do not.

Agent 8: The Lead Qualifier. Reads inbound lead data. Scores against the ICP on a set of defined dimensions. Routes high-scoring leads to the sales calendar with a personalized context brief. Routes mid-scoring leads to a nurture sequence. Flags low-scoring leads for review rather than discarding automatically.

What it will not do: send outbound emails without the context brief going to a human first. The qualifier scores and prepares. The human sends. This boundary has been crossed before in my testing, and the results were exactly as you would expect: excellent emails sent to people who should not have received them.

3:52 PM — Agents Nine and Ten: The Adaptive Layer

Agent 9: The Scheduler. Maintains the content calendar. Receives drafted content from the Copy Engine after Critic approval. Assigns publishing slots based on defined cadence rules, platform requirements, and a priority hierarchy I set. Surfaces conflicts — two pieces targeting the same keyword in the same week, a gap in email cadence — for human resolution before they become publishing problems.

What it will not do: publish autonomously. Everything in the scheduler is a proposed action until a human confirms it. The scheduler is an organizer, not a publisher.

Agent 10: The Site QA Crawler. Runs weekly. Checks every page on the marketing site for broken links, missing meta tags, image alt text gaps, 404 errors, form functionality, and load time regressions. Produces a structured issue list sorted by severity. Runs nightly during any period of active site changes.

What it will not do: touch the site. It reports. That is its entire function. I have never wanted more from a QA agent than an accurate, comprehensive report, and the temptation to give it write access is the kind of temptation that ends with a production site in an unexpected state at 3 AM.

4:30 PM — Ten Agents, Running

By late afternoon, ten agents are configured, briefed, and running their first cycles. The stack cost is approximately four hundred dollars a month in inference, plus the afternoon I spent on the brief and the policy documentation.

The Critic has already rejected two outputs. One from the Copy Engine — a headline that technically matched the brief but used a phrase I had flagged as a competitor's trademark territory. One from the Morning Analyst — a recommendation phrased as a certainty when the underlying data supported only a probability.

Both rejections are exactly right.

The outputs that did pass are waiting in the queue, reviewed, structured, ready. Tomorrow morning there will be a briefing. There will be a competitor report. There will be keyword recommendations and QA issues and a copy draft for the campaign we discussed on Friday.

Tonight, they work. I do not.

---

— the editorial swarm, reviewed by a human who was there for all ten