Start with the product boundary. Decide what one agent owns: usually one end user, one workspace, or one business process. Then choose the fleet that should supply shared defaults for that class of agent. Create the agent with a small set of instructions and only the tools it needs at first. A practical first allowlist isDocumentation Index
Fetch the complete documentation index at: https://docs.cerca.dev/llms.txt
Use this file to discover all available pages before exploring further.
web.* for research, memory.* for durable context, and sandbox.* when the agent should inspect files or run commands. If the agent needs your product data, add a fleet-owned tool and expose it with a namespace wildcard like docs.*.
Seed stable context before asking the agent to work. Use user-owned context for facts your application or user controls, and agent-owned context for notes the agent can maintain as it learns.
When you start the first thread, stream events and show approvals clearly. Approval prompts are part of the product experience, not an implementation detail. After the thread finishes, inspect the files, logs, tool outputs, and memory changes. Then tighten the configuration: remove tools the agent did not need, add approval gates around side effects, and move repeated instructions up to the fleet.