The 4 components of every AI agent: tools, context, skills, and model

Most people think AI agents are complicated. They're not. Every AI agent — from a basic customer support bot to a fully autonomous digital coworker — is built from the same four ingredients: tools, context, skills, and model. Get the four right and you build systems that feel like magic. Get them wrong and you end up with an expensive chatbot that hallucinates and breaks.

This piece breaks down each component, why it matters, and what separates the version that works from the version that doesn't.

Quick reference

• Tools — how the agent acts on the world
• Context — what the agent knows
• Skills — how the agent decides what to do
• Model — which AI handles each step

---

1. Tools — how the agent acts on the world

Tools are how an agent reaches outside its own conversation and changes something. There are two flavors that matter.

Persistent connections to your systems of record. Your CRM. Your inbox. Your calendar. Your database. Your project tracker. Your internal documentation. These are the systems where work actually lives. An agent without authenticated, durable access to them can describe what should happen but cannot make it happen.

Built-in capabilities. Web search, file creation, image generation, code execution, document reading. These are the general-purpose moves the model itself can perform — the digital equivalent of opening a browser tab or scratch pad.

An agent without good tools is like a brilliant employee with no laptop, no logins, and no phone. The reasoning is there. The ability to act is not.

The standard most teams will see in production over the next year is MCP — the Model Context Protocol — an open standard introduced by Anthropic in late 2024 for connecting AI assistants to data sources and tools. If a vendor offers an MCP server, you can plug it into any compliant agent runtime. This is what's driving the recent shift from "build a custom integration per tool" to "compose your agent from existing connectors."

The practical test. When the agent decides to act, can it actually act? Can it move a deal stage in your CRM, write a calendar invite, post a Slack message, or commit a file? If the answer is "no, it just produces text recommending those actions," you have a chatbot, not an agent.

---

2. Context — what the agent knows

Context is the agent's memory. This is where most agent projects quietly fail.

A model with no context is brilliant but amnesiac: it can reason about anything you put in front of it and remembers none of it afterward. Most teams skip context engineering, deploy a model wrapped in a basic prompt, and wonder why their "agent" needs every situation re-explained from scratch.

Good context is two things at once.

Indexed, searchable memory the agent can read and update. Not a single growing transcript. Structured files the agent owns: who the user is, what the team's preferences are, what was decided in the last conversation, what conventions are in force for this customer or this project. The agent reads what it needs at the start of a task and writes back what is worth remembering at the end.

Session management that knows when to retain, compact, or reset. Long conversations don't fit in a model's context window. The agent must continuously decide what to keep, what to summarize, and what to drop — without losing the load-bearing details. This is engineering, not prompting.

The signal that context is working: the agent improves over time without anyone retraining the underlying model. It learns your team's vocabulary, your customers' edge cases, your preferences for how things should be written. A well-designed memory system turns a one-trick tool into a colleague that genuinely knows your business.

The signal that context is broken: the agent feels like a stranger every time. You catch yourself explaining the same project setup, the same customer history, the same constraints. If you ever say "as I told you last time" — there is no last time. The agent never remembered.

---

3. Skills — how the agent decides what to do

Skills are how the agent knows what to do in a specific situation. Think of them as formalized playbooks.

When a competent human employee encounters a familiar situation — a customer complaint, a refund request, a sales lead, a bug report — they don't reason from first principles. They follow a pattern. Step one. Step two. Decision point. Step three. The pattern itself is institutional knowledge. The pattern is what makes the response repeatable, fast, and consistent across different employees on different days.

Agents work the same way. A skill is a documented procedure the agent loads when it recognizes a relevant situation. It describes:

• When to invoke this pattern
• What tools to use, in what order
• What the success criteria look like
• What to do if the situation deviates from the standard path

Anthropic formalized this idea in late 2025 with its Agent Skills feature: a portable format where a skill is a directory containing a markdown specification and any supporting files. The agent reads the skill, follows the procedure, and returns. Skills are composable, version-controlled, and shareable across teams.

Why this matters in practice. The best agents aren't just smart. They're repeatable. An agent improvising from scratch on every refund request produces ten different outcomes. An agent following a refund skill produces the same outcome every time — and when the policy changes, you update the skill once.

Skills are also how you scale judgment without scaling people. The senior employee who knows exactly how to handle the complicated case can encode that knowledge into a skill, and the agent applies it consistently across every instance.

---

4. Model — which AI handles each step

Model is the last ingredient, and the most misunderstood. People obsess over "which model is best." That's the wrong question.

Speed, cost, and intelligence are not properties of a model. They are outputs of two inputs: the context you give the model and the model you pick for the specific task.

A few practical truths:

• Smaller, faster models with great context outperform frontier models with none. A Haiku-class model fed exactly the right three paragraphs of context will beat an Opus-class model fed an unstructured wall of text. Specificity matters more than raw intelligence.
• An agent is rarely one model. Production agents typically route different sub-tasks to different models: a small model for classification, a mid-tier model for the bulk of conversational work, and a frontier model reserved for the hardest reasoning steps. This is sometimes called a model router.
• Cost compounds. A workflow that runs a frontier model on every step can cost 50–100x more than the same workflow with appropriate routing. That is the difference between an agent that's affordable to deploy at scale and one that's only economical for the occasional VIP customer.

The right question is never "which model is best." It's "which model is right for this specific step, given the context I have to give it, and the latency and cost I can afford here?" Match the model to the job. Don't default to the biggest one.

---

The four together

These four components — tools, context, skills, model — are all an agent is.

The agents that feel like magic aren't doing something fundamentally different. They have all four dialed in at the same time:

• The tools they need, on the systems they need to act on
• The context to recognize the situation and remember what's been decided
• The skills to execute the pattern that's known to work
• The right model for each step, with the right cost-to-quality tradeoff

Skip one and the others can't compensate. A perfect model with no tools is a chatbot. Perfect tools with no context is brittle. Perfect skills with the wrong model is slow and expensive. Perfect context with no skills is endlessly improvisational and inconsistent.

The teams building agents that actually work in production aren't picking the fanciest model. They're investing in all four at once. That's what Velanir does — we hire, configure, and operate digital coworkers with all four components engineered together, so the agent walks in already knowing your business and ready to do the work.