Create Custom Agents in Minutes via Natural Language or Skills

Use the Claude console's "new managed agent" option or trigger the TypeScript Claude SDK skill (v0.10+) with /claude-api managed-agents-onboarding to build agents from English descriptions. For a medical agent example: instruct Claude to read Markdown files from a private GitHub repo, interpret as a doctor, and respond via Slack. It auto-recommends read-only tools (read, glob, grep—no write/edit/bash to avoid mutations), generates system prompts, suggests Sonnet model for cost savings over Opus, and creates setup.ts (deploys env/agent on Anthropic infra) and runtime app.ts (Slack bot integration via .env credentials). Run setup.ts to get env/agent IDs visible in console; run app.ts for Slack queries like "Is calamari safe?"—agent greps repo, analyzes allergies (e.g., shrimp cross-reactivity causing itchy tongue/swelling), responds concisely after prompt tweaks for human tone. Edit agents in console UI: update prompts, models, tools across versions without redeploying.

This skips SDK mastery or VPS setup, hosting code/tools on Anthropic's sandboxed infra for instant production use cases like customer Q&A from knowledge bases or scheduled Slack research.

Harness-Session-Orchestrator Delivers Fault-Tolerant Scaling and Security

Agents run on three core components for reliability: Harness (stateless router, Claude-powered "hands") executes tool calls, MCPs, code in sandbox, or file edits—separating logic from state prevents leaks. Session (append-only memory) logs all harness activity; if harness fails, Orchestrator spins a new one that replays logs to resume exactly. Supports unlimited models/environments at scale.

Security via runtime credentials: Vault stores API keys (e.g., GitHub, Slack, weather MCP, even Anthropic's own)—injected per-tool at execution, invisible to model/harness. No .env exposure risks. Pre-built MCPs for Notion/Slack; custom via code. Beats OpenClaw (tinkerer Linux-like: self-host hardware/models/security) by handling infra like Apple—secure, hands-off for production.

API-Only Pricing and Integration Limits Demand Trade-off Evaluation

Billed per Claude API tokens (Pro/Team subs ignored) + $0.08/session-hour (active only; idle free). Sonnet/Haiku keeps costs low vs Opus. Lacks broad channels (e.g., Telegram/Discord/WhatsApp)—code custom integrations or use OpenClaw for openness. UI suits no-coders; code for edge cases. Ideal for secure prototypes to prod (e.g., medical data queries), but pair with self-hosted for flexibility.