Customizing an agent

The copied factory is deliberately small. Edit it directly for product-specific behavior and pass runtime capabilities through its config when they should remain replaceable.

Tools and MCP

import { webSearch } from '@agentskit/tools'
import { createMcpClient, toolsFromMcpClient } from '@agentskit/tools/mcp'

const mcp = await createMcpClient(/* transport */)
const agent = createMyAgent({
  adapter,
  tools: [webSearch(), ...(await toolsFromMcpClient(mcp))],
  onConfirm: (call) => approve(call),
})

Use onConfirm for tools that write data, spend money, publish content, or trigger external actions.

RAG and memory

const agent = createMyAgent({
  adapter,
  retriever: knowledgeBase,
  memory: conversationMemory,
})

Retrieval supplies grounded context for a run. Memory preserves conversation state. Treat both as untrusted input: enforce tenant boundaries and keep prompt-injection defenses in the skill.

Observability

const agent = createMyAgent({
  adapter,
  observers: [tracer, auditLog],
})

Record latency, model usage, tool calls, errors, and approval decisions. Redact personal data and secrets before sending traces to third-party systems.

Delegation

Agents expose .asHandle() for supervisor and swarm topologies. Keep delegation depth and step limits bounded, and propagate cancellation signals through every child run.