Agentic Engineering

Core principles

1. Tools over prompts: Give agents tools to ACT, not just instructions to THINK. A web search tool is more reliable than asking the LLM to recall facts.
2. Context is everything: Agent quality = f(context quality). Give precise, relevant context. Remove noise.
3. Eval-driven: Define what "correct" means BEFORE building. If you can't eval it, you can't improve it.
4. Fail gracefully: Agents will fail. Design for recovery (retries, fallbacks, human escalation).

Tool design

• Tools should have clear, specific descriptions (the LLM reads them to decide when to use each)
• One tool per action (don't combine "search AND create" into one tool)
• Return structured data, not prose
• Include error information in tool responses (don't silently fail)

Context management

• Limit context to what's relevant for the current step
• Use progressive disclosure: metadata first, full content on demand
• For long documents: chunk and retrieve, don't dump the whole thing in context
• Track token usage. Context windows are finite.

Multi-agent patterns

• Router: One agent classifies the task, routes to specialist agents
• Supervisor: One agent delegates tasks and reviews results from worker agents
• Pipeline: Sequential handoff between specialized agents (research → write → review)
• Swarm: Multiple agents with shared state, each working independently

Eval patterns

• Define pass/fail criteria BEFORE implementation
• Use LangSmith datasets for regression testing
• Test tool calling accuracy (does it call the right tool with the right args?)
• Test output quality (does the response answer the question?)
• Test edge cases (what happens with empty input? malformed data?)