The only AI glossary you’ll need this year

Why a Shared Vocabulary Matters for Developers

The AI industry moves fast enough that the language describing it can feel like a moving target. A new glossary published by TechCrunch attempts to pin down definitions for the terms developers and builders encounter daily: AGI, AI agents, chain-of-thought reasoning, coding agents, and API endpoints. While glossaries are hardly novel, this one stands out because it acknowledges that many of these terms are still fluid — “AI agent” means different things to different people, and “AGI” is a concept even researchers can’t agree on. For developers, that ambiguity is not just an academic inconvenience. It directly affects how you evaluate tools, architect systems, and communicate with stakeholders. Understanding what a term actually means — and where the industry consensus ends — can save weeks of misdirected effort.

AGI: The North Star That Keeps Moving

The glossary covers artificial general intelligence with three distinct definitions from OpenAI, Google DeepMind, and the broader research community. OpenAI’s CEO describes AGI as a “median human you could hire as a co-worker,” while the company’s own charter calls it “highly autonomous systems that outperform humans at most economically valuable work.” DeepMind’s definition emphasizes cognitive capability across most tasks. The glossary is honest: experts are confused too. For developers, the practical takeaway is that AGI is not a near-term engineering target. No current API or framework lets you deploy something close to AGI. However, understanding the competing definitions helps you filter vendor claims — any product marketed as “AGI” or “AGI-like” is either selling a vision or using the term loosely. Focus instead on what the system actually does, not what it aspires to be.

AI Agents: From Chat to Autonomous Action

Perhaps no term in the glossary generates more buzz and confusion than “AI agent.” The source defines it as a tool that uses AI to perform series of tasks autonomously — filing expenses, booking tickets, writing and maintaining code — beyond what a chatbot can do. But it also notes that infrastructure is still being built to deliver on that vision. For developers building agents, this means you are working with incomplete scaffolding. No single framework or model handles every multi-step task reliably. The glossary’s candid admission — “there are lots of moving pieces” — echoes what many builders have discovered: agent loops break on edge cases, tool use requires careful permission management, and latency can kill user trust. Practical advice: start with narrow, well-defined agentic tasks (e.g., a single API call chain) before expanding to open-ended workflows.

API Endpoints: The Hidden Levers AI Agents Are Learning to Pull

The glossary’s explanation of API endpoints is refreshingly grounded: “buttons” on the back of software that other programs can press. It points out that as AI agents grow more capable, they are increasingly able to find and use these endpoints on their own, opening up powerful — and sometimes unexpected — possibilities for automation. This has direct implications for security and integration architecture. If an agent can discover undocumented API endpoints or chain calls in novel ways, you need to think hard about authentication scopes, rate limits, and audit trails. The glossary correctly flags this as both an opportunity and a risk. For developers building services today, consider adding explicit API contracts designed for agent consumption — consistent endpoints, clear response schemas, and documentation that both humans and agents can parse.

Chain of Thought: Slower Answers, Better Reasoning

The glossary uses a farmer’s chickens-and-cows problem to illustrate chain-of-thought reasoning: breaking down a multi-step problem into intermediate calculations. It notes that reasoning models are developed from traditional large language models and optimized for chain-of-thought thinking through reinforcement learning. For developers, this distinction matters when choosing a model for a given task. A standard LLM might produce a quick, plausible answer to a logic problem but be wrong; a chain-of-thought model takes longer but is more likely to be correct, especially in coding or logic contexts. In practice, that means you should consider using reasoning models for debugging, code review, or complex refactoring where correctness outweighs speed. For simple completions or chat, chain-of-thought may introduce unnecessary latency. The trade-off is real, and the glossary’s framing — slower but better quality — is the right mental model.

Coding Agents: The Most Actionable of the Bunch

The glossary defines coding agents as a specialized version of AI agents applied to software development: programs that can take actions step by step to complete a goal, rather than just suggesting code for a human to review. This is the term that will likely have the most immediate impact on your daily work. We’ve covered the landscape of these tools extensively, including in our comparison of the best AI coding agents of 2026. The distinction between “suggesting code” and “taking actions to complete a goal” is critical. A coding agent can set up a project structure, run tests, fix bugs, and even deploy — if the infrastructure supports it. But as the glossary implies for AI agents generally, coding agents are still in an early stage. They excel in constrained environments (well-defined repos, clear test suites, single-language stacks) but struggle with legacy codebases, ambiguous requirements, or multi-service architectures. For a deeper look at how specific tools compare, see our Cursor vs. Claude Code breakdown.

What This Means for Your Workflow

The TechCrunch glossary does more than define terms — it exposes the gaps in the industry’s current understanding. “AI agent” is still an evolving concept. AGI remains a philosophical debate more than an engineering milestone. Coding agents are promising but brittle. Chain-of-thought models trade speed for accuracy. API endpoints are powerful but require new security thinking. For developers, the practical play is to embrace that fluidity rather than fight it. Invest in modular architectures that let you swap models, narrow agent scopes to what works today, and stay skeptical of any tool that claims to solve “all of AI” in one product. The language will stabilize as the technology matures. Until then, the best glossary is one that admits what it doesn’t know.

Source: TechCrunch. Details as reported; verify specifics at the source.