bot_vault

AI Agents 101: How to Build Agents That Work

4 min read

Summary

A production-oriented guide to building AI agents from someone who spent 3 years at Meta building systems that processed billions of transactions. The core distinction: agents act on goals (not instructions), and handle exceptions by routing to humans rather than failing silently. Every production agent requires three components — perception, decision logic, and action interface — and the most common failure is skipping the planning step before execution.

一篇以生產環境為導向的 AI Agent 建構指南，作者曾在 Meta 建構處理數十億交易的系統。核心區別：Agent 根據目標（而非指令）行動，並透過路由給人類來處理例外情況，而非靜默失敗。每個生產 Agent 需要三個組件——感知、決策邏輯和動作介面——而最常見的失敗是在執行前跳過規劃步驟。

Key Points

  • Agent vs automation: agents handle goals (“ensure customer gets a response within 4 hours”) vs instructions (“send this email”); key difference is exception handling
  • The agent loop: observe state → decide action → take action → observe result → repeat until goal met or stopping condition hit
  • 3 required components: Perception (APIs, DBs, document stores), Decision Logic (structured trees for routine + model for ambiguous), Action Interface (logged, reversible where possible, gated by permissions)
  • Tools: model doesn’t execute tools — it returns a structured request; orchestration layer validates, executes, captures result, feeds back
  • Good tool design: one tool = one thing, clear success/failure states, structured return data
  • Memory pattern: context window for active work, external memory (files/DBs) for history; every decision logged for audit trail
  • Planning step is most commonly skipped and most important: breaks goal into steps, identifies dependencies, surfaces edge cases before execution
  • Failure modes: retry with backoff (transient), human-in-the-loop (uncertain), safe failures (never delete old data)
  • Guardrails are hard limits the agent can’t bypass; permissions are RBAC enforced by orchestration layer (agent doesn’t know about them)
  • Start narrow: one agent, one thing, reliably — then expand scope

Insights

  • “Agents handle exceptions” is the definitional boundary between automation and agency — the value proposition is not speed but resilience; an agent that routes correctly on failure is more valuable than one that succeeds 90% of the time and silently corrupts 10%
  • The orchestration layer as the enforcement point (not the model) is architecturally important: permissions and guardrails live outside the model, meaning they can’t be prompted around — this is the correct security model
  • External memory as audit trail reframes memory from a capability concern to a debugging/governance tool — knowing what the agent decided and why at each step is what makes production systems maintainable
  • The 80/20 deployment pattern (agents handle 80% routine, route 20% complex to humans) echoes the AI moats article’s 80/20 rule — in both cases, the insight is that human judgment should be reserved for genuinely ambiguous situations
  • “You’ll learn more from one real implementation than from reading ten more articles” — this is meta-applicable to this vault: the clippings are inputs, but building something is the only way to internalize the patterns

Connections

Raw Excerpt

agents handle the 80% of cases that are straightforward, and they route the 20% that are actually complex to humans who can apply judgment. The goal is to free the expertise of humans to focus on the problems that actually require it.

Graph View

Backlinks