Planning: What Separates a Chatbot From an Autonomous Agent

Posted on Oct 25 • Originally published at agentic-affairs.com

Every vendor is selling you an "agent" these days. Strip the marketing and the real dividing line is embarrassingly simple: a chatbot answers, an agent finishes. One responds to a prompt. The other carries a goal across steps, watches what happens, and decides what to do next without you holding its hand.

That's the distinction this piece nails down — and it's one the industry keeps blurring on purpose, because "agent" sells better than "chatbot with a loop."

A system can write a beautiful plan and still not be an agent. If a human executes every step and decides when to continue, the planning is real and the autonomy is not. The test that matters is not "does it form a plan?" but "who forms the plan, who executes it, and who decides whether to continue, replan, or stop?" Anything less is a chatbot wearing an agent costume.

Perceive, plan, act, check, repeat. Anthropic's version is a while-loop wrapping LLM and tool calls; OpenAI's is reason, call tool, observe, answer. The piece is honest about what that costs: across 500 SWE-bench coding tasks, agentic coding burned roughly 3,500x the tokens of a single-round reasoning task — $1.857 per task against $0.016 — because accumulated context gets re-sent every round. Agents don't just cost more. They cost more because of the loop itself, and nobody selling you an agent platform puts that number on the landing page.

Even the best frontier agent completed only 30.3% of TheAgentCompany's 175 long-horizon workplace tasks, at $4.20 a pop, with failures that included getting stuck in loops or aimlessly exploring the environment. METR's 80%-success horizon is roughly five times shorter than its 50% horizon. Planning makes multi-step work possible. It does not make it reliable — and anyone demoing a flawless agent run is showing you the one run that worked.

One question, one answer, nothing to decompose, observe, or recover from? A chatbot is the right tool, and reaching for an agent is burning tokens to feel futuristic. Reach for planning when the goal is multi-step, open-ended, or needs recovery from failure — filing expenses, investigating a bug across a codebase. The discipline is knowing which side of the line your task sits on before you pick the architecture.

The full piece walks through the perceive-plan-act-check loop, ReAct and Reflexion, the LangGraph recovery patterns, and the three classic planning failures — over-planning, brittle plans, and plans that hallucinate tools that don't exist. Read it here: Planning: What Separates a Chatbot From an Autonomous Agent

Originally published on Agentic Affairs ( https://agentic-affairs.com )