Native LLM Capabilities

Built-in features that enhance your Agent’s reasoning, memory, and ability to interact with users and data.

Capabilities are native features of the underlying LLM that extend what your Agent can do—beyond just processing instructions and calling actions.

Enable them with the capabilities array on defineAgent. Each entry is a capability slug (or a { slug, config } object for ones that take options):

agents/researcher.ts

export const researcher = defineAgent("researcher", {
  connector: openai,
  languageModel: "gpt-4o",
  systemPrompt: "Research accounts and produce a shareable brief.",
  capabilities: ["webSearch", "memory", "document", "suggestedActions"],
});

The full set of slugs: webSearch, memory, document (Canvas), suggestedActions, action, file, model, documentationSearch, sandbox, and context. The sections below explain the most common ones and when to enable them.

file and model are also enabled automatically when files or models are attached to the Agent, so you rarely list those by hand. file covers listing, reading, searching and analysing files — analysing runs code over them in a sandbox, which is what the separate codeInterpreter slug used to do.

---

Available capabilities

Memory

Memory allows Agents to retain and recall information across interactions, enabling personalization and continuity.

Memory type	Scope	Use case
User Memory	Persists per end-user	Stores preferences, past interactions, and personal context for tailored responses
Workspace Memory	Shared across all Agents	Organizational knowledge like company policies, ICP definitions, or team guidelines
Agent Memory	Current execution only	Short-term context including previous steps and action outputs within a single run

Use Workspace Memory for information that all Agents should know—like your ICP criteria or product positioning.

---

Web Search

Gives Agents access to real-time information from the internet.

When to enable:

• Account research requiring current market data
• Validating company information that may have changed
• Finding recent news, funding announcements, or leadership changes

Example query:

“What is the latest funding round for Acme Corp?”

---

Canvas

Enables Agents to generate structured, shareable digital assets. The capability slug is document (surfaced as Canvas in the web app).

Supported outputs:

• Spreadsheets and CSV files
• Text documents
• Formatted reports

Example use case: A user asks: “Give me a list of CROs in San Francisco.”

The Agent researches, compiles the data, and delivers a formatted spreadsheet—ready to download or share.

---

Suggested Actions

A proactive assistant that recommends the next best step after completing a task.

How it works:

1. Agent completes the requested task
2. Analyzes what logical next steps might help the user
3. Offers actionable suggestions

Example: After retrieving a list of contacts, the Agent suggests:

“Would you like me to add these contacts to Sequence XX?” “Should I enrich these contacts with email addresses?”

Suggested Actions keep workflows moving and help users discover capabilities they might not have thought to request.

---

Actions

action lets the Agent search and run any action in the workspace, instead of only the ones you attached in uses. It is the same four steps Cargo’s MCP server exposes, running inside the Agent: search the catalog, read the action’s input schema, resolve any field that needs a picked id, then execute — over one record or a batch.

agents/operator.ts

export const operator = defineAgent("operator", {
  connector: openai,
  languageModel: "gpt-4o",
  systemPrompt: "Handle inbound GTM requests end to end.",
  capabilities: ["action"],

  // Still pinned as their own tools, and preferred when they fit
  uses: [enrich, hunter.actions.findEmail],
});

Enabling it is the grant: execution is not limited to the actions you attached. Attached actions stay first-class tools — their inputs are known up front and their configuration is fixed — and the Agent is told to prefer one of those when it matches the job.

When to enable:

• Open-ended operator Agents, where you cannot predict which integration a request will need
• Exploratory work, where pinning every candidate action ahead of time is impractical

Actions consume credits. The Agent is instructed to quote the cost from the catalog before running, and to pilot a sample before a large batch, but keep this capability off Agents that should only ever do the one job you pinned.

---

Models

model gives the Agent your data models: list them, read their columns, search records, and run read-only SQL.

It has two widths on the same slug:

• Enabled on its own — every model in the workspace, read-only.
• With models attached to the Agent — only those, with each attachment’s filter, column selection and row limit re-applied on every call. This is also the only way to allow writes, by attaching a model that is not read-only.

Attaching a model enables the capability automatically, so you rarely list model by hand unless you want the whole-workspace read.

---

Enabling capabilities

Toggle capabilities on or off based on your Agent’s purpose:

Agent type	Recommended capabilities
Research Agent	Memory, Web Search, Canvas
Qualification Agent	Memory, Suggested Actions
CRM Updater	Memory only
Operator Agent	Actions, Models
Conversational Assistant	All capabilities

More capabilities = more complexity. Only enable what your Agent actually needs to avoid unpredictable behavior.