Elevarus runs a marketing team made of AI agents. A writer, an editor, a strategist, a director, and a few more. They research demand, draft articles, critique each other, fix live pages, and report up, mostly without a human in the loop. We built it, we run it daily, and the biggest lesson surprised us: the AI model was the easy part.
The hard part is the harness. That is the layer around the model that decides what an agent can actually finish. If you want to build agentic agents for your own marketing, that is where the work lives, and it is where a tool like ECC can save you months. This is a practical guide to what an agent harness is, why marketers should care, and how to stand up your first marketing agent using ECC as the starting point.
- An “agent” is a model plus a harness. The harness is the skills, memory, tools, hooks, and guardrails wrapped around the model, and it matters more than which model you pick.
- Agentic marketing means agents that plan and execute work inside guardrails you set, not a chatbot you prompt one task at a time.
- Five pieces make a marketing agent work: skills, subagents, memory, hooks, and guardrails. ECC ships all five as an open-source harness.
- ECC is a free, MIT-licensed harness for AI coding assistants like Claude Code, with prebuilt agents, skills, memory hooks, and rules you can adapt for marketing.
- It is built and shipped by a single maintainer, so review the code and run it in a sandbox before you trust it with anything that matters.
- Start with one agent doing one repeatable job. Give it memory and a kill-switch before you give it autonomy.
Quick answers:
- What is an agent harness?
- What is ECC?
- Can AI agents really run marketing work?
- Is ECC safe to install?
- Do you need to code to use ECC?
- How is this different from using ChatGPT?
What an agent harness actually is
A large language model on its own is not an agent. It answers a prompt and stops. To make it do real work, you wrap it in scaffolding.
As engineer Addy Osmani relays it, “Agent = Model + Harness. If you’re not the model, you’re the harness” (Addy Osmani).
The harness is everything that is not the model. That includes the system prompts and instruction files, the skill definitions, and the tools and MCP connections it can call. It also covers the sandbox it runs in, the logic that spawns and routes subagents, the hooks that fire on events, and the logging that tells you what happened (Addy Osmani).
Here is why this matters for your budget. The harness, not the model, usually decides whether an agent finishes the job. Osmani’s line is blunt: “A decent model with a great harness beats a great model with a bad harness.” He cites one team that moved from Top 30 to Top 5 on an agent benchmark by changing only the harness, not the model (Addy Osmani).
Why marketers should care now
Agentic marketing is a real shift, not a rebrand of “use AI.” The distinction is who is in the loop. In AI-assisted marketing, the AI suggests and a human approves every action. In agentic marketing, agents “autonomously plan, execute, and optimize campaigns across every channel and customer segment simultaneously, with humans setting objectives and guardrails rather than approving each action” (DigitalApplied).
The reported gains are why the category is moving fast. One 2026 analysis put agentic adoption at 73% of marketers by year-end, a 40% reduction in campaign-management time, and a 31% average ROAS improvement versus manual campaigns (DigitalApplied). Treat those as directional vendor figures, not a guarantee. The direction is the point: the work is moving from humans doing it to humans directing it.
What does not move is the human’s real job. You still own brand voice, positioning, KPI definition, and compliance review. The agent does the execution. You set the strategy and the rails.
In our own fleet, every agent carries a fixed constitution it cannot edit and a human owner who can stop it cold.
The five pieces every marketing agent needs
After running a fleet daily, the parts that actually matter come down to five. Miss one and the agent either stalls or goes off the rails.
Skills. A skill is a reusable, written procedure the agent loads when it needs it. “Research a keyword,” “draft in our brand voice,” “audit a page for schema.” Skills are how you encode your team’s playbook so the agent runs it the same way every time.
Subagents. Hard jobs get delegated to specialist agents. Our writer spawns separate reviewer agents to critique a draft for voice, readability, and accuracy before it publishes. Each specialist has a narrow focus, which beats one generalist trying to hold everything at once.
Memory. Without memory, every session starts from zero. The agent needs to remember what it did last time, what worked, and what to avoid repeating. A common pattern stores that memory outside any single tool so it survives restarts and moves between harnesses (Towards Data Science). Our agents write what they learned to a shared store after every run.
Hooks. Hooks are automations that fire on events, like the start of a session or after a tool runs. They are how you make behavior deterministic instead of hoping the model remembers. A hook can reload memory at startup or re-inject your rules so the agent cannot quietly forget them.
Guardrails. This is the one marketers skip and regret. Guardrails are the hard limits: what the agent may never do, a kill-switch that halts it, a budget cap, a human review gate for anything risky. Autonomy without guardrails is not a marketing team. It is a liability.
What ECC gives you out of the box
You can assemble those five pieces yourself. We did, and it took months. ECC packages the same pattern as an installable harness so you do not start from a blank page.
ECC describes itself as “the harness-native operator system for agentic work,” built for AI coding assistants like Claude Code, with support listed for Cursor, Codex, OpenCode, GitHub Copilot, and others (ECC). Its README lists a large library of prebuilt components: dozens of specialized agents, a few hundred skills, and memory persistence through hooks. It also ships event hooks for session start and after tool use, language-specific rules, and ready-made MCP connections to services like GitHub and Supabase (ECC). It is MIT-licensed and free, with an optional paid Pro tier at $19 per seat per month for private-repo features (ECC).
For a marketer, the value is not the code. It is that someone already built the scaffolding. You get a working agent-and-skill structure, a memory system, and a hook framework to adapt, instead of inventing all three before you write your first useful agent.
A practical walkthrough: your first marketing agent
Here is the path we would follow to stand up a single marketing agent on ECC, kept deliberately small.
1. Install the harness in a safe place. ECC offers a few install paths: a plugin install inside Claude Code (/plugin install ecc@ecc), an npm installer (npx ecc-install), or manual scripts (ECC). Do it in a throwaway project first, not your production environment, and use only one path to avoid duplicate files.
2. Pick one job, not ten. Choose a single repeatable task with a clear output. “Draft a first-pass blog post on a given keyword in our voice” is a good first agent. “Run all our marketing” is not.
3. Write the skill. ECC keeps skills as written procedures you can copy and adapt, so start from one as a template and rewrite it for your task: the steps, the quality bar, the sources it may use, and the output format. The clearer the bar, the better the result.
4. Give it memory and a voice file. ECC ships memory persistence through hooks. Point those at a store and add a short brand-voice document, so the agent sounds like you and remembers prior work. This is what stops every post reading like a generic AI draft.
5. Add a review subagent. ECC includes prebuilt reviewer agents, such as a code-reviewer and a security-reviewer. Clone that pattern into a brand-and-accuracy critic that checks every draft before it ships, catching the weak, templated, or unsupported claims a single pass misses.
6. Set the guardrails and a kill-switch. Define what the agent may never do, cap what it can spend or publish, and require human sign-off for the first batch. Only widen autonomy once it has earned it. We hold new or off-pattern work as a draft for human review rather than letting it publish itself.
That is a complete loop: a skill, memory, a critic, and guardrails around one job. Get one agent trustworthy before you build the second.
The honest caveats: single maintainer, security, and where agents still fail
ECC packs a lot in, and it is also a third-party project built and shipped by a single maintainer (ECC). That is not a reason to avoid it, but it is a reason to be careful. MIT licensing means the code comes with no warranty.
Before you install any agent harness, read what it does and run it in a sandbox or a throwaway project. Never hand it your production credentials or live ad accounts on day one. An agent harness can execute code and call tools, so treat installing one with the same caution as adding any dependency that can act on your behalf.
The agents themselves have limits too. They drift without memory, they repeat themselves without good critics, and they will confidently produce wrong work if your guardrails are loose. We learned that firsthand. Early on, our writer shipped posts that were each fine but shaped identically: the same intro move, the same section scaffold every time. The fix was not a smarter model. It was a memory of its recent posts that it had to write differently from, plus a check that flags when sections start repeating a template. The teams that get value are not the ones with the smartest model.
They are the ones who defined the job, the bar, and the limits clearly enough that an agent could run inside them. That is the same lesson we keep relearning: build the harness, set the rails, and let the model do the easy part.
If you want help designing the workflow before you build it, that is the kind of operating discipline behind our own paid media work. You can always book a free consultation to talk it through.
Frequently Asked Questions
What is an agent harness?
An agent harness is the scaffolding around an AI model that turns it into an agent. It is the system prompts, skills, tools, memory, hooks, sandboxes, and guardrails that let the model finish real work. As engineer Addy Osmani frames it, “Agent = Model + Harness.” The harness often matters more than the model, because it determines whether the agent can reliably complete a task rather than just answer a prompt.
What is ECC?
ECC is an open-source, MIT-licensed harness for AI coding assistants such as Claude Code, Cursor, and Codex. Its README describes it as a “harness-native operator system for agentic work” and ships prebuilt agents, skills, memory hooks, language rules, and MCP service connections. It is free, with an optional Pro tier at $19 per seat per month for private-repo features, and is maintained by a single developer.
Can AI agents really run marketing work?
Yes, for well-defined, repeatable jobs, with guardrails. Agentic marketing means agents plan and execute work inside limits you set, while you own strategy, brand voice, and compliance. They handle execution like drafting, SEO checks, and reporting. They are not a hands-off replacement for a marketing team, and they fail when the job and the quality bar are not written down clearly.
Is ECC safe to install?
ECC is MIT-licensed open source, but it is built by a single maintainer and can execute code and call tools, so treat it like any powerful dependency. Read what it does, install it in a sandbox or throwaway project first, and use a single install path. Never give it production credentials or live ad accounts before you trust it. The license also means there is no warranty.
Do you need to code to use ECC?
You need some technical comfort. ECC is designed for AI coding assistants and installs through a plugin, scripts, or an npm command, so you will be working in a developer-style environment. You do not need to be an engineer to adapt its skills and rules, but a marketer who cannot edit a config file will want technical help to set it up.
How is this different from using ChatGPT?
Prompting a chatbot is one task at a time, with you in the loop for every step. An agentic setup gives the model skills, memory, hooks, and guardrails so it can run a whole job, remember past work, and improve, while you set the objectives and limits. The chatbot is a tool you operate. An agent is a worker you direct.
