Build Claude agents end-to-end

The Claude API / Agents SDK gives direct access to Claude models with maximum flexibility — the developer controls the whole stack. It supports stateless API calls, multi-turn agents that loop on tool calls, extended thinking for complex reasoning, and MCP integration to reach any external system.

Unlike Glean (indexed enterprise search) or OpenAI (managed assistants), Claude is unconstrained on data sources: any API, database, or service is reachable through defined tool schemas.

AESOP runs the generic API-agent pipeline — fitness check, gap analysis, workflow design, then instruction writing — and emits:

• System prompt — the agent’s instructions and guardrails.
• Tool schemas — function-calling definitions for the external systems it touches.
• MCP config — connections to any MCP servers the workflow needs.

Because the developer owns the full stack, the build is the most open-ended of the targets: no fixed platform contract, just a well-scoped agent.

Deployment means standing up the custom backend (or wiring a stateless API call). The deployed agent is then run through AESOP’s Evaluate stage and scored for quality against the PRD.