The pet agent dilemma

Every day on my social feed I see people talking about their fleet of agents in Claude. Rich - the lead enrichment agent, Lil' buddy - the marketing assistant, and on it goes.

I love to see people offloading their anti-to do lists, but truth be told, you don't need a specialized agent for every little thing you do. And it might actually hold you back.

Why I only use one agent

I don't build agents for every workflow. Instead - no surprises here - I use Adapt.

A single universal agent like Adapt (or a general-purpose agent like Claude Cowork or Manus) is able to replace most specialized knowledge work agents. Today's frontier models, equipped with tools, context, and compute, can do the work of your many specialized agents, including spinning up their own sub-agents and orchestrating work to achieve a goal.

Equipped with agent Skills, they can go even further to repeatably perform any workflow with high levels of accuracy.

Limiting scope limits learning - and unlearning

Most of us don't know what super-intelligence feels like at work, and we never will unless we experience it directly. We learn fastest when we interact with models directly, open our minds to what might be possible now, and simply play.

When you limit the scope of an agent, unless it's an incredibly specific workflow where you need the agent to have memory and improve over time, you're probably going to miss what the model itself is actually capable of.

For example, Claude Fable 5 was released earlier this week. If your lead enrichment agent has very specific instructions for how it operates, you might never even realize that Fable 5 can do insane levels of de-anonymization and linking across your data in half the time at twice the quality of Opus.

I've been playing with Fable 5 for the last 48 hours and I can definitively say that workflows I just couldn't do before are now possible.

An agent anecdote

In the spring of 2025, I briefly explored building an agentic system for content, primarily to avoid n8n and Gumloop for building content workflows.

I ended up never kicking off the project and joined Adapt a few months later. Using Adapt solved everything I planned to do with the planned system within 2 months. (And I haven't even thought about using n8n or Gumloop since.)

I think a lot about how if I had invested in that project I would have likely been stuck maintaining outdated workflows for my agent fleet instead of experimenting with the latest models and understanding what I no longer needed as part of that system.

The opportunity cost of losing even one month of direct learning with the latest models is undeniably high. Businesses especially can't afford to let their employees fall behind because they're managing a fleet of specialized agents whose work is now obsolete.

There is and will be a time for specialized agents

I do believe specialized agents will run more and more of our workflows over time, as we enter the era of indisputable super-intelligence.

But more than that, I believe that companies that let their employees invest in this reality too early without support will miss out on understanding frontier capabilities that will become necessary to survive in the AI era.

No one wins when departments are operating in AI silos building and re-building the same agents for the same workflows. You'll quickly end up with 7 different competitive analysis agents relying on 3 different sets of context, too.

Elaine Zelby of Tofu says it well:

Centralizing agent sprawl with a company brain

To manage the inevitable sprawl Elaine mentions, companies are turning to building their own company brain. Companies like Ramp and Sentry have built and announced their own solutions.

Adapt exists, in part, because not every company has the resources and talent that Ramp and Sentry have to build their own solution. Most companies we talk to are working to make their products AI-native first, and their employees second. That doesn't negate the importance: the fastest way to working AI-native company-wide is to provide a central "company brain" and an agent that runs on it.

Adapt's company brain acts as a shared context layer made up of your tools, knowledge, and workflows, and gives everyone on your team a frontier agent to use wherever they work. It's a system with shared skills, shared workflows, and a collaborative work surface to learn from each other in.

As models improve, Adapt improves automatically, so your whole team's work stays frontier quality.

We're running a 60-day pilot program with uncapped usage for qualified business pilots, enough to get your company running with frontier AI. If that sounds interesting, set up a demo with our team or sign up and try it yourself with $100 free credits.