The Five Concepts That Separate Agentic Coding From Autocomplete

01 A Vocabulary Becomes a Curriculum

On October 12, IBM Technology - the education channel of a company that sells standardization - published a ten-minute explainer titled 5 Concepts to Know for Coding with AI Agents. By October 15 it had drawn roughly 122,000 views. When the education arm of IBM decides practitioners need a canonical five-item vocabulary, the vocabulary is crossing from community jargon into infrastructure.

Curricula are how practices cross from demo to discipline: what cannot be named cannot be taught, hired for, audited, or procured. The concepts circulating through enterprise training - context engineering, tool use, verification, guardrails, orchestration - are less a syllabus than a fossil record. Each one exists because a specific way of failing proved common enough to need a name.

02 From Prediction to Action

The line the vocabulary draws is the line between prediction and action. Autocomplete completes a token in a file a human is already editing; the human remains the actor. An AI agent, in the standard definition, is a program that can make decisions, pursue goals, and take actions independently. Agentic coding moves the actor: the system plans, edits across files, runs commands, calls services.

That transfer of agency is what makes new vocabulary necessary. Every delegated action is code that ships without a human watching it get written. The failure modes of that arrangement were not hypothesized; they were discovered in real repositories, at production cost. The five concepts are the controls that got built over the holes.

03 Every Concept Is a Scar

Read the list as incident history. Context engineering - the deliberate curation of what enters a model's working context - exists because agents fail when they cannot see the relevant code, and confidently modify the wrong file. Tool use discipline exists because a malformed or hallucinated call does not fail safely; it corrupts state. Verification exists because generated code arrives confident and wrong at a rate that broke autocomplete-era review assumptions. Guardrails exist because autonomy without boundaries acts outside them. Orchestration exists because multi-step work drifts: a plan that made sense at step one is unrecognizable by step nine.

The mapping is the analysis. No concept on the list is about extracting more impressive output from a model; every one is about constraining a failure mode somebody already shipped. Marketing adjectives describe upside - every term here is downstream of a scar.

04 The Common Thread: Supervision

What the five share is a single rearrangement of the developer's job. In the autocomplete era the human wrote and the model predicted; the skills were fluency and taste. In the agentic era the human specifies, reviews, and verifies; the core skills are context curation, boundary-setting, and audit - the job description of a supervisor of non-human engineers. The vocabulary formed top-down: enterprises needed terms for managers to scope work and auditors to demand artifacts.

05 The Maturity Ladder

The vocabulary also encodes a progression, and it is worth being precise about rungs. Autocomplete: the human writes, the model predicts; blast radius: the current file. Chat assistance: the human describes, the model drafts, the human pastes; blast radius: what the human accepts. Agentic coding: the system acts across the repository and the human reviews after the fact. Governed agent pipelines: the system acts inside explicit controls - and the blast radius is bounded by policy rather than by vigilance.

06 What a Standard Curriculum Says About Production

The contrast with the vibe-coding era is the clearest measure of the shift. Vibe coding, as the term settled, meant describing a task in natural language and accepting the generated code without thorough review, guiding changes by results and follow-up prompts. It was fast and useful for prototypes - and incompatible with production, because its core move was the removal of review.

The new vocabulary is the correction, concept by concept. Verification reinserts the check that vibe coding removed. Guardrails reinsert the boundary. Context engineering reinserts knowledge of the codebase. Orchestration reinserts the plan. A practice gets a standardized curriculum when it becomes production infrastructure - when failures cost more than education. The five concepts exist because enterprises started paying for the failures.

07 Limits and What to Watch

The limits deserve equal billing. Naming a control is not enforcing it; a workforce can learn the vocabulary and skip the discipline; compliance theater is real wherever the curriculum arrives faster than the tooling. The concepts are also unevenly mature - verification inherits decades of testing practice, while orchestration has almost none.

Watch for the vocabulary getting teeth. Whether procurement language starts requiring verification and guardrail artifacts the way it now requires security review. A vocabulary that stops changing is the sign a discipline has settled; this one is still moving.