Skill, Workflow, or Agent? One Question Decides

The most expensive habit in AI engineering right now isn't picking the wrong model — it's building an agent for a job that a fixed sequence would handle better. Anthropic's own engineering guidance is blunt about this: agentic systems "trade latency and cost for better task performance", and their first recommendation is to find the simplest solution possible. Autonomy is something you pay for. Most tasks don't need it.

The three tiers — skill, workflow, agent — are really one variable: who decides the steps. Once you see that, the buzzwords collapse into a decision you can make in ten seconds.

Skill: nobody runs steps — it's packaged knowledge

A skill is know-how you write down once: a folder with a SKILL.md holding instructions plus optional scripts and reference files. Anthropic shipped Agent Skills on October 16, 2025, and the engineering behind them is the interesting part — progressive disclosure. Only the skill's name and description sit in the system prompt; the body loads when the task matches, and bundled files are read only if needed. That makes the reference material you can attach "effectively unbounded" without burning your context window.

Reach for a skill when you keep re-explaining the same procedure — your PR review checklist, your report format, your team's release conventions. Rakuten put a number on the payoff after adopting them: work that used to take a day now takes an hour.

Workflow: you decide the steps

Anthropic defines workflows as systems where models and tools are orchestrated "through predefined code paths" — and names five patterns: prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer. In Claude Code, the everyday version is a slash command: a markdown file (.claude/commands/deploy.md, or a skill you trigger as /deploy) that fires your exact sequence on demand.

This is the right tier for anything repetitive with a known shape: same input type, same ordered steps, same output format. It's deterministic, testable, and cheap, because the model executes your path instead of reasoning about which path to take.

Agent: the model decides the steps

Agents, per Anthropic, "dynamically direct their own processes and tool usage." You hand over a goal and tools, not a script. In Claude Code that's a subagent — it runs in its own separate context window, with its own tool permissions, and you can even pin it to a cheaper model like Haiku for grunt research. The cost of that power is reasoning at every turn: more tokens, more latency, more variance between runs.

The one question

Before building anything, ask: could I write the exact steps down right now, before the task starts?

• Yes, and it's mostly "how we do X here" knowledge → skill.
• Yes, an ordered sequence that must run identically every time → workflow.
• No — what happens next depends on what comes back → agent.

Why this matters beyond your terminal

"Agentic AI" is on every international job post now, and most candidates can only recite hype. Being the person in a system design round who says "that's not an agent problem — a routed workflow does it deterministically at a fraction of the cost" is a senior signal, and it lands especially well when you're interviewing in English for a company abroad. This taxonomy is cheap to learn and it compounds.

Today: take the one task you've explained to Claude more than twice this month and freeze it as .claude/skills/<name>/SKILL.md in your repo. That's tier one of the ladder — and now you know exactly when it earns a promotion.

Sources

• Anthropic Engineering — Building effective agents
• Anthropic — Introducing Agent Skills
• Anthropic Engineering — Equipping agents for the real world with Agent Skills
• Claude Code Docs — Create custom subagents