Agentic Is Not Magic. It's Management.

I've had this conversation too many times recently. I bring up building a brain or giving an AI a framework. The other person cuts in. "Is that agentic?"

Every time. As if saying the word closed the loop.

Most people are using "agentic" or "agents" as magic words. Accio agents. The words have stopped describing and started substituting. Say them, get the magic.

Anthropic's definition is the one I keep coming back to:

That's the whole definition. Decides for itself.

Agentic AI is not magic. It's the ability to build really good employees.

Issue 1 was the brain. Issue 2 was memory and context. Agents are the team that works inside that brain.

The skill to build this well isn't technical. It's leadership. And if you've ever built a team, hired a specialist, or set expectations for a direct report, you already have it.

So the question is how. How do you translate leading people into building agents?

I've built teams for twenty years. Watched hundreds of leaders do the job. The best ones all give their people three things. Trust. Context. Agency. That's the bar. It doesn't change because the hire is an agent.

Trust. Whether it understands what you actually want. The judgment you don't need to double-check.

Context. The memory the agent draws on. The brand, the history, the standards. Everything a new hire gets on day one.

Agency. Whether it's allowed to act on its own. When it escalates instead of acting.

Three simple things. The job of management, human or not.

That's the frame. Now the team.

One person can run marketing. Not by hiring one AI to do everything. By hiring four AI specialists. The section will explain.

You run marketing at a mid-size beauty brand. That's who you are. Your job is taste and instinct. Keeping the brand and having a feel for what lands. The AI's job is to build the rest.

The rule from Issue 1 holds. AI earns its keep by doing one of three things.

Efficiency.
Reach.
New capabilities.

Every specialist on your team should be pulling at least one of those levers. If an agent is doing something you could do alone, it doesn't belong on the team.

Consumer listening. [Lever: reach.] Reads every review, social mention, return reason, and CS ticket. Clusters by SKU, claim, and ingredient. Writes a weekly signal brief. No human gets through that volume at that frequency.

Opportunity scanner. [Lever: reach + new capability.] Watches competitor launches, pricing, category trends, and loyalty patterns in parallel. Surfaces patterns a human wouldn't connect across sources. Hands you a weekly shortlist. You pick what to pursue.

Ad variant generator. [Lever: reach.] Takes an approved hook and drafts fifty copy and concept variations across channels. Static, paid social, video scripts. You pick the winners. It's a volume engine, not a creative director.

Performance analyst. [Lever: efficiency + new capability.] Reads post-campaign data. CTR, CVR, AOV, channel mix. Writes the retro and proposes the next test before the current one is finished. The retro takes minutes, not days. The next experiment is queued before the dust settles.

Four agents. One team. One person running it.

Taste is what you bring. Scale is what AI brings. The team works when neither side tries to do the other's job.

An agent is a markdown file. You can read it, edit it, and version it. Writing it well is the whole job.

Two file names you'll see: claude.md (Claude's tools) and agents.md (the emerging cross-tool standard used by Cursor, Codex, and most others). Same idea either way. It lives in the specialist's folder.

Best practices are converging. The four principles below get you to a working first specialist. It won't be perfect. It'll get better every time you correct it. That's the whole point.

Onboard it like a new hire. The instruction file is the onboarding packet. What your business does, who your customer is, what you ship, what you won't. Focus on what's unique to your brand: your voice, your claims rules, your no-go list. Don't explain what a headline is; the model already knows. Don't paste your full brand guidelines either. Put them in /context/brand-voice.md and reference them. Aim for under 200 lines. If a human couldn't do the job from your instructions, an agent can't either.

Be testable, not tasteful. "Use our brand voice" is useless. "Avoid the word 'clean.' Cap adjectives at three per headline. No puns in video hooks" is testable. Vague rules get ignored. Specific ones get followed.

Name what it can't do. Guardrails matter as much as permissions. "Does not publish. Does not pick winners. Does not spend budget." Saying what the agent can't touch is what keeps it useful instead of dangerous.

Give it a way to check its own work. The single highest-leverage move: tell the agent how to verify its own output before handing it back. "Check each variant against the claims standards file. Flag any variant that breaks a rule." Without verification, you are the only feedback loop.

An agent without feedback loops is a stamp. Paste in input, get output. No growth. Three moves turn a stamp into an employee.

Two files, two jobs. The instruction file is what you write. Rules, role, process. A second file, memory.md, is what the agent writes about your feedback. Preferences, corrections, patterns. Keep them separate.

Close the correction loop in the instructions. Add one rule near the top of the instruction file: "When I correct you or you learn something new, update memory.md with the correction and the context it applied to." That single line turns every correction into a permanent change. The VP kills a pun-heavy batch once. The agent writes it down. Next batch, no puns. Forever.

Skills compound. When you notice the agent rebuilding the same process from scratch every time, write it down once as a skill. /skills/headline-variant-protocol.md. /skills/claims-safe-language-check.md. Skills are SOPs for AI. Explain something once, never explain it again.

Here's what the ad variant generator's instruction file looks like when the principles are applied.

Short. Every section does one thing. Every path updates in one place. Every correction compounds.

The file is what you hire. The feedback is what you manage.

We haven't covered the whole team. A real marketing org has more than four specialists. There's a sequencer keeping launches honest. A brand reviewer gating every output. And there's the workflow that connects them all. Who hands off to whom, when, and what happens when things break.

That's orchestration. How agents hand off to each other. How they run in sequence. How they escalate. When one agent's output becomes another's input.

Issue 4 is where we go there. Today was hiring. Next is managing.

Ethan Mollick said it plainly:

The team will be imperfect. Hire your first agent anyway.

— Nick