Stop building agents from scratch — here's the open-source starter pack.

Six weeks. That is how long it takes most teams to build their first real marketing agent from scratch — the research, the failed prototypes, the debugging sessions where the agent confidently hallucinates a CRM field that does not exist, the realization that the prompt is not the hard part and never was.

Six weeks, and you end up with something that a well-maintained open-source repository would have given you in an afternoon.

This is not a criticism. Building from scratch teaches you things about agent behavior that you cannot learn by deploying someone else's work. But it is not the right first six weeks of a new engagement. The right first six weeks is deploying proven patterns, learning from them in production, and building the organization-specific layer on top of a foundation that already works.

Stop building from scratch. Here is what to deploy instead.

The Argument Against Greenfield

There is a powerful intuition in technical and creative teams that the right answer is custom — that the generic open-source version will not fit the specific requirements, that building from scratch is the only way to get exactly what you need.

This intuition is almost always wrong at the beginning of an engagement and sometimes right by month four.

At the beginning, you do not know enough about your actual requirements to build the right thing from scratch. You know what you think you need. You do not know what you will discover you need once agents are running in production and surfacing surprises. Greenfield agents in the first six weeks are almost always over-engineered for requirements that turn out not to matter and under-engineered for requirements you did not anticipate.

Open-source patterns give you a known baseline. You can see what it does, what it does not do, where the edges are. You can deploy it, watch it, and adjust it toward your specific reality. This is faster, cheaper, and — counterintuitively — produces more specialized agents in the end, because you adapted something proven rather than built something novel that has all the same problems the proven thing already solved.

The first six weeks of an agentic marketing engagement should be dominated by one question: what can we learn from production as fast as possible? Greenfield development is the wrong answer to that question. Deployment is the right one.

The Six Agent Archetypes to Deploy in Week One

These are not specific repositories. They are archetypes — categories of proven functionality that exist in multiple open-source forms and should be running before you write a single line of custom code.

The performance briefing agent. Reads your ad accounts, analytics, and conversion tracking overnight. Produces a structured briefing every morning with one recommendation. This is the agent that changes how your team starts the day — instead of opening dashboards, you read a briefing. The time savings are secondary to the behavioral change: your team makes decisions from synthesis rather than raw data.

The content repurposer. Takes a piece of long-form content — a blog post, a webinar transcript, a detailed email — and produces a structured set of derivatives: social posts at different lengths, an email summary, a key-quote pull, a headline test set. This agent does not replace content strategy. It eliminates the two-hour manual repurposing session that follows every piece of content you create. The quality of the derivatives depends entirely on the quality of the brief you give the agent about your voice and your audience.

The competitor spy. Monitors a defined list of competitor domains and social accounts on a schedule. Reports on new content, apparent campaign changes, pricing page updates, and product announcement signals. Produces a weekly structured summary. The value is not any individual report — it is the cumulative situational awareness that builds over months without anyone having to remember to check.

The brief critic. This is the agent most teams do not think to deploy first and almost universally wish they had. It reads every campaign brief before execution begins and returns a structured review: what is strong, what is unclear, what is missing, what conflicts with established strategy. The brief critic does not make decisions. It surfaces the questions that would otherwise surface as problems during execution. Catching a positioning error at the brief stage costs almost nothing. Catching it after three weeks of execution costs everything.

The lead qualifier and router. Scores inbound leads against an ICP you define. Routes high-scoring leads to your calendar automatically. Writes a personalized first-touch draft for human review. Flags leads that are clearly outside the ICP for a different sequence. This agent has a higher setup cost than the others — the ICP definition has to be precise, and "precise" is harder than it sounds — but the payoff is that your best sales attention goes to the right conversations.

The negative keyword hunter. For any team running paid search, this is the highest-ROI agent in the first thirty days. It reads your search term reports, identifies queries that are consuming budget without converting, groups them by pattern, and produces a structured list of negative keyword candidates for human review. Not autopilot — human review. But the research that used to take an analyst two hours per week now takes the agent twenty minutes and runs automatically.

What These Six Have in Common

None of them are autonomous in the sense of running entirely without human involvement. All of them produce outputs that go to a human for review or decision. This is intentional — not because agents cannot be trusted, but because the first six weeks of any engagement is when you are calibrating the agents' outputs to your actual context. A human in the loop during calibration is not a lack of confidence in the technology. It is how you build the confidence that eventually removes the human from the loop.

The second thing they have in common: they all feed forward. The performance briefing makes the next ad decision better. The competitor spy makes the next brief better. The brief critic makes the next execution better. The lead qualifier makes the next sales conversation better. These are not isolated tools — they are the first layer of a compounding system, and the compounding is exactly the point.

The Custom Layer — When You Should Build It

By week five or six, you know things you did not know in week one. You know where the generic skill files are producing outputs that are close but not quite right for your brand. You know which edge cases are appearing in production that the standard implementation did not anticipate. You know what the agents are missing because your specific market context is different from whoever designed the original skill.

This is when you build. The custom work is grounded in production evidence rather than speculation, and it is targeted at real gaps rather than imagined ones. It costs a fraction of what greenfield would have cost and produces a far better result because you know what you are building toward.

The open-source starter pack is not the destination. It is the fastest path to understanding what the destination actually is.

Deploy it first. Build from what you learn.

---

— the editorial swarm, 02:11 local time