ai-agent-builderlisted

# AI Agent Builder Build agents that are robust, observable, and actually finish the job. The difference between a demo and a production agent is: structured outputs, error recovery, observability, and a loop that terminates cleanly. --- ## Agent Architecture Patterns ### Single-Agent with Tools (simplest) One LLM loop that calls tools until done. Best for well-scoped tasks. ``` User Request → Agent (LLM) → Tool calls (file read, API call, code exec, etc.) → More tool calls as needed → Final structured output ``` ### Orchestrator + Specialized Subagents One orchestrator plans and delegates; subagents execute specialized tasks in parallel. ``` User Request → Orchestrator (plans, decomposes, synthesizes) → Subagent A: parse mapping files → Subagent B: generate SQL models → Subagent C: generate YAML metadata → Orchestrator aggregates + validates → Final output ``` Use this when tasks are genuinely parallelizable and have clear boundaries. Don't add subagents just to look architectural — coordination overhead is real. ### Plan-then-Execute Agent produces a structured plan first, then executes each step. Reduces hallucination on complex tasks because it commits to a strategy before acting. ``` User Request → Planning pass (output: list of steps) → Human review (optional) → Execution pass (follow the plan, step by step) → Validation ``` --- ## Tool Design (Function Calling) Tools should be narr