Model, context, skill, tool, agent, squad

The first stumble after leaving regular chat is calling everything an agent. You switch the model and expect it to know the company. You write a long instruction and expect it to become a tool. You connect a tool and expect it to solve lack of context. The frustration comes from there.

What I do today is separate six things before designing any automation: model, context, skill, tool, agent, and squad. Each one answers a different question.

Model

The model is what reasons. It is the part that reads the question, combines patterns, writes text, and decides the next step. When I switch the model, I am switching the quality and type of reasoning: more or less care, more or less ability to follow a long chain, more or less discipline when using tools.

When do I need this? When the task needs better reasoning or a different way of working. Difficult architecture, careful review, writing with constraints. Switching models does not teach anything about me. A better model still knows nothing about my projects if I do not give it context.

Context

Context is what the system knows without me explaining it again. It is not the whole conversation. Conversation evaporates and carries noise. Good context is a file: a written decision, customer profile, product rule, glossary, history that someone can open again.

When do I need this? When the answer depends on me, the company, or a project. If I ask for a proposal, sales plan, or review without context, the model returns an average of the internet. The best model still writes generic output when the base is generic.

Skill

A skill is a procedure the agent learned to recognize and follow. It is not memory. It is a way of working. If every time I review a plan I check risks, blockers, and minimum tests, that becomes a skill. The value is not having to explain the ritual again.

When do I need this? When the task repeats with similar steps. A good skill does not try to store everything; it says how to act, what to verify, and when to stop. If the problem is lack of information, a skill does not solve it. It only executes a procedure better.

Tool

A tool, or MCP, is the ability to act outside the model itself. Read a database, search the web, open a document, operate another application. Without a tool, the model only talks. It can write a perfect plan and still change nothing in the world.

When do I need this? When the task requires external state or external action. Check the calendar, query a base, create a card, read a file that is not in context. The tool does not think for you; it gives hands to the reasoning.

Agent

An agent is a role with context, tools, and an objective. It has a name because you want to call that role again: PR reviewer, researcher, analyst, editor. The name keeps you from rebuilding the same combination every time.

When do I need this? When recurring work needs the same posture and the same access set. If everything changes with every request, maybe it is only a conversation with a good instruction. An agent starts to pay off when the role becomes stable.

Squad

A squad is more than one agent working on the same task, with distinct roles. One researches, another implements, another reviews. The idea is to separate responsibilities that fight when they live in the same head.

When do I need this? When the task has independent parts or needs real tension between roles. For small things, a squad only increases coordination. The step exists, but it is not the first one.

The question that closes almost every diagnosis is: what am I trying to do right now: give context, follow a procedure, use a tool, or call a specialized role? It avoids the common mistake of asking the model for something that should be in a file.

To start on the right step, the work is the same as always: write the context file before anything else.

Where it breaks

The ladder invites you to climb too early. Almost everyone who wants to build a squad has not written the context file yet. So they create several agents, each with a nice name, and all of them return different versions of the same generic answer.

The real cost is coordination before substance. Without context, a skill becomes an empty ritual, a tool becomes a button without judgment, an agent becomes a persona, and a squad becomes a meeting. Start with the boring step. Write the base before asking the system to act as a team.