Tools & Skills

Fruxon provides two mechanisms for equipping AI agents with capabilities: Tools and Skills. Both give agents the ability to interact with external systems, but they differ in how and when those capabilities become available during a conversation.

Tools

Tools are individual actions an agent can perform — calling an API, querying a database, sending a message, etc. When you attach tools directly to a prompt step, the agent has immediate, always-on access to every tool in the list. The LLM receives the full schema (name, description, parameters) of every attached tool on every request.

How Tools Work

1. You attach tools to a prompt step in the agent workflow
2. On every LLM call, all tool schemas are included in the context
3. The agent can call any tool at any time during the conversation

When to Use Tools

• The agent needs a small, focused set of capabilities (1–5 tools)
• Every tool is relevant to every conversation the agent handles
• You want the simplest setup with no activation logic
• The agent's role is narrow and well-defined (e.g., a retrieval agent that always searches and summarizes)

Skills

Skills are reusable capability bundles that combine tools, instructions, and resources into a single activatable package. Unlike tools, skills use a progressive disclosure model — the agent only sees a brief description of each skill until it decides to activate one. Upon activation, the skill's full instructions, tools, and resources are injected into the conversation.

How Skills Work (Progressive Disclosure)

1. Discovery — The agent sees only the skill name and description via the activate_skill tool. This is lightweight — just enough for the agent to decide if it needs the skill.
2. Activation — When the agent calls activate_skill, the skill's detailed instructions (procedural knowledge) are injected into the conversation as a tool response.
3. Execution — The skill's tools become available in the agent's active tool set, and any linked resources (documents, knowledge bases) become accessible.

When to Use Skills

• The agent serves multiple domains or workflows (e.g., a customer support agent that handles billing, technical issues, and order management as separate concerns)
• You have many tools (10+) and want to avoid overwhelming the LLM's context window with irrelevant tool schemas
• You want to include procedural instructions that are only relevant when a specific capability is needed
• You want to share a standardized set of capabilities across multiple agents (e.g., a "Salesforce CRM" skill used by both your sales agent and support agent)
• You want the agent to make intelligent decisions about which capabilities to engage based on the conversation

Comparison

How Skills Reduce Context Window Usage

Every tool attached to a prompt step consumes context tokens — the LLM needs to see the tool's name, description, and full parameter schema to know how to use it. For agents with many integrations, this can consume a significant portion of the context window before the conversation even starts.

Skills solve this by replacing N tool schemas with a single short description. For example, instead of sending 8 Salesforce tool schemas (create lead, update opportunity, search contacts, log activity, etc.), the agent sees:

Salesforce CRM — Activate when the user needs to create, update, or search CRM records, log activities, or manage sales pipeline.

Only when the agent determines it needs CRM capabilities does it activate the skill, at which point the 8 tool schemas and CRM-specific instructions are injected.

Skill Anatomy

A skill consists of four parts:

1. Display Name & ID — Human-readable name and unique identifier for API reference
2. Description — A concise explanation of when the agent should activate this skill. This is what the LLM sees before activation, so it should clearly describe the skill's purpose and trigger conditions.
3. Instructions — Detailed procedural knowledge in Markdown. Injected when the skill is activated. This can include step-by-step procedures, business rules, response templates, escalation criteria, etc.
4. Tools — The set of integration tools this skill requires. Each tool can be bound to an agent-specific integration configuration (credentials/API keys) when the skill is attached to a prompt step.

Configuring Skills in Agent Studio

When you add a skill to a prompt step:

1. The skill appears in the Skills section of the step configuration
2. Expand the skill card to see its required tools
3. For each tool that requires authentication, select the appropriate integration configuration (credentials)
4. The skill's description is visible in the card; click the instructions icon to preview the full instructions in Markdown

If a tool is missing an integration configuration, an error indicator appears on the skill card. All tools must be properly configured for the skill to function at runtime.

Best Practices

Skill Design

• Write descriptions as clear trigger conditions: "Activate when the user asks about..." rather than just listing capabilities
• Keep instructions focused and actionable — the agent reads them at activation time, so they should be directly useful for the task at hand
• Group tools by domain or workflow, not by integration. A "Process Refund" skill might use tools from both Stripe and your internal API.

Choosing Between Tools and Skills

• Start with tools for simple agents. If you find yourself adding 8+ tools or writing complex system prompts with conditional logic ("if the user asks about X, use tools A and B; if they ask about Y, use tools C and D"), refactor into skills.
• Use skills when different conversations require different subsets of your total tool set.
• Use tools when every tool is relevant to every conversation.

Mixing Tools and Skills

You can use both on the same prompt step. Attach universally-needed tools (like a knowledge base search) directly, and package domain-specific capabilities as skills. The agent will have direct access to attached tools, plus the ability to activate skills when needed.

Next Steps

• Agent Studio — Visual workflow builder
• Creating Agents — Build your first agent
• AI Connections — Connect external tools