There’s plenty of debate about what exactly an “agent” is in our space. Here’s my take: an agent is a system that uses an LLM + tool calling + reasoning to iterate on an objective until completion. In short, it autonomously decides what to do next.

Assuming AI is benign or aligned, one huge problem we’ll face as society is the gap between AI being good enough to put folks out of jobs and not quite good enough to make up for it yet (scientific breakthroughs).

Well yea, the point of it having to exist in the first place and who created it is the issue I believe. Depending on how many tools / servers you have and the task you’re doing, could be less or more tokens. Just my observation lately.

MCP is a barrier for AI. You don’t need MCP servers to hit that API, just let your LLMs build its own tools.

I would pay monies for this!.

Why are there no good desktop MCP client / hosts? Sure there's claude desktop, but I want to use more models. Its baffling to me. I just tried 10 of them, all bad. Been holding off, but I think I'll just have to buy Raycast AI and use the MCP tool for that.

RT @petergyang: Let's get our definitions straight:. AI workflow = AI follows a pre-defined path. AI agent = AI decides the best path. Very….

If you’re using MCP just to wrap tools, you don’t need MCP. We’re slicing resources on a tool level when they should be sliced on a responsibility level imo. Also, MCP will be redundant soon, as AI won’t need us to write its tools. Would love to see more focus on applying AI.

RT @FrontendMasters: ▲ Next.js Fundamentals v4 by @scotups is out now!.👉 Build a real-world app & level up your Ne….

The debate on what constitutes an agent will continue, but I’m convinced that this spectrum; from scripted flows to fully autonomous systems; will shape the future of AI-driven automation. What are your thoughts.

What we’re seeing is more than just a new set of models, it’s a paradigm shift. The dynamic interplay between LLMs, tool calling, and free reasoning redefines what “autonomy” means in our systems.

This shift isn’t just academic. It’s driving real company value. By embracing these advanced reasoning models, businesses are reimagining automation, enabling agents that adapt on the fly and augment human productivity in truly innovative ways.

On the evaluation front, traditional evals have focused on aspects like tool selection accuracy. But as agents become more holistic, simulation-based evaluations might better capture their dynamic, real-world performance. Food for thought.

With these models, companies are integrating more sophisticated reasoning capabilities into their agents, achieving massive improvements in accuracy and robustness. The trade-off? A hit on speed and costs, which is often a fair compromise (for now).

Enter reasoning models like o1, o3 mini, and Deep Seek. These breakthroughs deliver “free” but not cheap reasoning, enabling real-time planning that’s critical for agents to iterate on complex objectives.

In practice, most systems lie somewhere in between. The degree of agency is dictated by how much decision-making power you grant the model—whether it’s simply following a set path or dynamically choosing its next action.

I see agency as a spectrum. On one end, you have pre-scripted workflows—where every decision is preordained. On the other, fully autonomous agents that leverage environmental resources (like databases or even monetization layers) to drive outcomes.

At its core, an agent or agentic system combines three pillars: an LLM for natural language understanding, tool calling to extend capabilities, and integrated reasoning to navigate tasks. This mix lets agents plan, pivot, and execute with a degree of autonomy.

RT @holtbt: Branching preview environments, now shipping in the @vercel + @neondatabase native integration! . Congrats to the Vercel team….

Trying to understand if @PlanetScale changed the design of their site for devs or for AI crawlers. Prob both. Works for them though.