agent computer apiAgent Computers APIInbound WebhooksOutbound WebhooksAPI OverviewAgent Computer InternalTasks API
agent computersAgent ComputerAgent HarnessAuthenticationSkills
agent contextAgent Context SkillAgent Context APIAgent ContextData RefreshCSV FilesJSON & JSONL FilesParquet FilesContext Server InternalsAzure Blob StorageDatabricksGoogle Cloud StorageMySQLPostgreSQLAmazon S3Snowflake
agent securityAccess ControlAgent Shield

Agent Context Skill

Every agent in your organization can query your connected data sources out of the box. This happens through the Agent Context skill — a built-in skill that is automatically installed on every Agent Computer when a task starts.

There is no manual step. You don't need to tell the agent to "use the data." If you've connected sources in the dashboard, your agents already have access.

How It Works

When you create a task, three things happen automatically:

  1. The skill is installed. The context-lake CLI is pre-installed on every Agent Computer during task setup. This gives the agent three commands: connect, catalog, and query.

  2. The system prompt tells the agent. Every agent's system prompt includes instructions to use the context-lake skill when the task involves structured data, business insights, or business logic. The agent knows to try Agent Context first for questions about your organization's data.

  3. The agent connects on demand. When the agent runs context-lake connect, the system provisions (or resumes) your organization's Agent Context engine, generates a scoped access token, and returns the query endpoints. This happens in seconds.

From the agent's perspective, it simply runs a CLI command and gets data back. All the infrastructure — VM provisioning, credential management, access control — is invisible.

The Three Commands

context-lake connect

Establishes a connection to your organization's Agent Context engine. Must be called first.

  • Wakes up the query engine if it was paused (auto-pauses to save costs when idle)
  • Returns HTTP and FlightSQL endpoints for querying
  • Generates a scoped JWT token based on the agent's user permissions

context-lake catalog

Lists all datasets and views the agent has access to, including table names, column names, and types.

Found 3 item(s) available to agents:

1. customers (dataset)
   Columns: id (Int64), name (Utf8), email (Utf8), created_at (Timestamp)

2. orders (dataset)
   Columns: id (Int64), customer_id (Int64), total (Float64), status (Utf8)

3. monthly_revenue (view)
   Columns: month (Utf8), revenue (Float64), order_count (Int64)

The agent uses this to understand what data is available before writing queries.

context-lake query

Executes a SQL query and returns results. Uses standard SQL (Apache DataFusion).

context-lake query 'SELECT name, email FROM "customers" WHERE created_at > '\''2025-01-01'\'' LIMIT 5'

Access Control

The agent only sees data it's allowed to see. Access is controlled by the same permissions you configure in the dashboard:

  • Shared datasets/views are accessible to all agents in the organization
  • Private datasets/views require explicit ACL grants to specific users
  • Admin users have access to all datasets and views
  • The agent's access token is scoped to exactly the tables it can read — nothing more

See Access Control for the full permission model.

← PreviousSkills
Next →Agent Context API