Why we built Fractl Agents.

Understanding AI's progression and use cases has become my full-time job at Fractl over the past two years. I spend 8–10 hours a day working with AI as it relates to the marketing work the agency does — in one capacity or another. I also have decent technical chops, having done full-stack development for at least part of my role over the past 12 years. I find the work fascinating, and the rapid changes drive me forward daily.

I also feel the disappointment.

AI is paradoxical: incredible capability paired with stubborn dumbness. It's sycophantic, non-deterministic in most use cases, and prone to what I think of as basin attractors — the gravitational pull every model has toward the safest, most generic answer. Even the best systems get pulled toward the boring. Prompting helps. Intuition for what works grows over time. But the frustrations don't disappear. Boring or obvious ideation. Hallucinations. Generic, uninspired research. If you've used AI seriously for marketing work, you've felt this.

AI can also be remarkable. It can teach you in the way you learn best, distill a hundred articles into the parts that matter, synthesize across messy inputs, and surface ideas you wouldn't have surfaced alone. It can be a real research partner.

These two paths look contradictory. They're not. AI is only as good as it's set up to be. It needs the tools, processes, workflows, and human-brain knowledge that real work depends on — the things that don't live inside the model and never will.

Anthropic is doing the best job I've seen here. Claude's agent product is genuinely good, and they keep improving its ability to use skills, tools, and connect components. But there's still real friction, even for technical people.

Am I picking the right skills? Is it using the MCPs and tools I gave it correctly? How do I help my non-technical colleagues do this same dance? And — most importantly — what are the best ways to actually leverage this glut of options on the work I do every day?

Tell me if this rings a bell.

I wanted AI to help me expand a list of pitch ideas for a client. The team had ideas, I had ideas, and I wanted to push past what we already knew — get a real sense of the client's industry and surface angles we hadn't considered.

ChatGPT did some good initial research. With the right prompt, the right model, and enough thinking time, it built a real picture of the industry — recent news, structural dynamics, the works. Then it gave me ideas. They weren't great. They didn't reflect the way Fractl evaluates content concepts, so I gave it our rubric. It scored its own ideas way too high. I wrote detailed feedback explaining why each concept didn't work. It tried again. The new ideas were marginally better. None were usable.

So I went back to ideating myself, with slightly more industry knowledge than I started with. Net win? Maybe.

What was the actual problem? AI isn't bad at this task — with the right setup, it can be genuinely good. But the right setup isn't a clever prompt. It's a system. Mine needed our evaluation rubric and an iterative loop, because first-pass ideas are never the best ones. It needed more than the base tools could give it — I'd already searched the internet, but I wanted podcast transcripts, YouTube content, Reddit threads. So I built tools. I gave it an SOP for how brainstorming actually happens at Fractl: the recursive steps, the analogical thinking, the way one analyst's wrong idea becomes another analyst's right one. I added sub-agents that evaluated outputs at each step. I borrowed skills other people had already shared. I added supplementary models from other providers, partly as collaborators and partly to compare which directions different models pulled toward.

After all of that — the rubric, the tools, the loop, the SOP, the sub-agents, the cross-model checks — I finally got ideas worth using.

What I'd actually built was a Playbook. A bundle of workflows, skills, knowledge, processes, and tools that let the AI operate like a real ideation analyst instead of a one-shot idea generator.

This is what it takes today to get AI outputs that match what a thoughtful human would produce. The model needs to be taught. It needs the abilities required to do the job. It needs to be guided, evaluated, and tested through real experimentation. There is no clever prompt that skips this work — there is only the work.

Which brings me to why this couldn't stay an internal tool.

My process for solving these issues wasn't optimal. Even with Claude making it easier to connect pieces, it was still complex, confusing, and often profoundly frustrating. Worse, it didn't fit how a real marketing team works. Many of the tasks we do require deterministic flow — the shape of the output has to be predictable every time. You can't ship a client a best guess pitch package. You need the same procedure to produce the same shape of output, run after run.

I needed a way to assemble Playbooks for the kinds of end-to-end tasks I'd been fighting through, without rebuilding the scaffolding every time. So I started naming the pieces.

It needed expert knowledge and standard operating procedures — the way a senior strategist thinks, written down, indexed, and made accessible to the agent. That became Experts in Fractl Agents.

It needed deterministic procedures we could count on and audit — the same scoring rubric, the same ranking step, the same quality checks, every time. That became Workflows.

It needed focused capabilities to augment what the model can do — composable moves any expert could pick up, not buried in a single agent's prompt. That became Skills.

It needed real connections to marketing data: APIs, integrations, search engines, podcast transcripts, news archives, social monitoring. Pre-wired and authenticated, not built from scratch every time. Those became Tools and integrations.

Put those four together and you get a Play — a chat session already shaped to do a specific marketing task end-to-end. Pick a Play, drop in the brief, the work runs. The output is the same shape every time. The pieces are swappable. New Plays can be assembled by anyone using the platform. And the outputs only get better as the models orchestrating them improve.

That's what Fractl Agents is, and that's why it exists.

If any of this rings a bell — if you've spent the last six months fighting one prompt and one rubric at a time — you have two choices. Keep building the setup yourself, one tool integration and SOP at a time. Or use what we've already built. Either way, the core insight stands: AI in marketing isn't going to be solved by better models. It's going to be solved by better setups around the models — Plays that bundle the knowledge, the procedures, the skills, and the tools that real work depends on.

That's what we shipped. That's what we'll keep shipping. And if the platform is useful to you, you'll build your own Plays on top of it, share them with people you trust, and the catalog will grow faster than any one team could on their own.

Output is the only thing that actually matters. The setup is what makes that output real.