🕹️ Agents & Workflows

Workflows vs agents (Anthropic’s “Building effective agents” taxonomy)

Pattern	What	Use when
Augmented LLM	LLM + retrieval + tools + memory	Base building block
Prompt chaining	Fixed sequence of steps	Decomposable tasks, with gates between steps
Routing	Classify, then send to a specialized path	Distinct input categories
Parallelization	Sectioning / voting	Speed or confidence
Orchestrator–workers	LLM breaks down work dynamically	Unpredictable subtasks
Evaluator–optimizer	Generate → critique → refine	Clear evaluation criteria
Autonomous agent	LLM in a loop with tools and environment feedback	Open-ended problems where some autonomy is acceptable

Start with the simplest thing that works. Add agency only when it measurably helps.

Tool design: clear names/descriptions, typed inputs, idempotent, least privilege, good error messages
Loop control: max steps, budgets (tokens/time/$), stop conditions
State & memory: short-term (thread), long-term (store), summarization
Human-in-the-loop for irreversible actions (payments, emails, deletes) ⭐
Durable execution: agents as long-running workflows (Temporal, LangGraph checkpoints); retries without double side effects
Multi-agent: supervisor, handoffs, A2A protocol for cross-agent interop (awareness)
Security: prompt injection via tool results, sandboxing code execution, excessive agency
Observability: trace every step, tool call, and token

Temporal (Go/Java SDKs): workflows, activities, retries, signals, versioning ⭐
Step Functions / Airflow / Camunda (awareness)

🟡 All 6 workflow patterns as Orbit workflow definitions
🔴 A ReAct agent from scratch in Go with budgets and loop detection
🔴 A durable agent: kill the worker mid-agent → resume from the checkpoint without repeating side effects
⚫ Temporal vs Orbit engine vs LangGraph for the same agent; write the ADR

Go deeper

🧩 AI & Agent Patterns · Distributed & Cloud Patterns