Every few months, the favorite AI keyword changes. The actual engineering progress doesn't stop — but the vocabulary you have to wade through to find it does a full reshuffle.
It's worth keeping a list. Partly because it's funny, and partly because pattern recognition makes you a better engineer. Each of the last four terms-of-the-moment started as a real, useful concept and then got stretched until it meant almost nothing.
Here's the 2024–2026 cycle.
2024: “agentic”
The starting word. “Agentic” existed in academic literature before, but it got crystallized by Anthropic's December 2024 piece Building Effective Agents, which drew a sharp line between workflows (orchestration is fixed) and agents (the model decides what to do next). That distinction is genuinely useful — we still use it in our own work.
The trouble was that within a few weeks, every LLM call with a system prompt was being called “agentic.” A summarization endpoint with a JSON schema? Agentic. A retry loop? Agentic. The word stopped doing useful filtering.
Early 2025: “scaffolding”
Once “agentic” got diluted, the conversation moved to how you build them — which created room for “scaffolding.” Hugging Face's smolagents framed it cleanly: scaffolding is the structure you add around an LLM to make it useful — planning intervals, tool organization, memory layers, multi-agent splits.
Real, important concept. Within months, “agent scaffolding” started showing up on slides where the “scaffold” was three prompts in a YAML file. The word got bigger; what it described stayed about the same.
Mid 2025: “harness”
Inherited from the eval world. EleutherAI's lm-evaluation-harness — the boringly excellent toolkit behind Hugging Face's Open LLM Leaderboard — uses “harness” to mean the test framework that wraps a model and runs it through standardized tasks. That's a real, specific thing.
In mid-2025, “harness” jumped from evals to agents. Suddenly every agent SDK had a “harness layer.” Sometimes that layer was thoughtful tool routing with proper observability. Sometimes it was a function called run_agent() that took a config object. Same word, two orders of magnitude apart in what it actually contained.
Now: “agent sandbox”
You can guess the pattern. The current term-du-jour is “agent sandbox.” There's a real engineering pattern here — running agent-generated code in an isolated environment with restricted permissions, scoped tools, and proper observability. Daytona, E2B, Modal, Vercel, and latest from Google and Anthropic Managed Agents are all serious infrastructure for this, with measurable differences (Daytona cold-starts in ~27ms; E2B uses Firecracker microVMs per session; Modal can hold a GPU inside the sandbox).
And then there's the “agent sandbox” that is, on inspection, a try/catch around an await llm.call(). Both will be labeled the same way in 2026 pitch decks.
Why this matters (briefly)
It's tempting to roll your eyes at the cycle and ignore it. Don't. Two things to keep in mind.
The original concept is almost always worth understanding. Underneath every overused buzzword there's an engineering pattern that someone built carefully and gave a precise definition to. Real agent sandboxing — Firecracker microVMs with scoped IAM and tool allowlists — is a powerful thing to know how to build. Don't let the dilution rob you of the actual content.
Pattern recognition keeps you grounded. Every cycle has the same shape: a precise term appears, gets adopted, gets stretched, gets replaced by the next precise term — and in nine months that one will be diluted too (remember MCP?). If you can spot which stage of the cycle a word is in, you can listen past the noise instead of being swept along by it.
A small bet
The 2026 H2 buzzword is going to be “agent runtime” or “agent kernel.” You can mostly tell which one is winning by watching the ratio of framework READMEs to marketing decks. When the decks start outnumbering the READMEs, the word is about six months from being replaced.
The work isn't in the word. It's always in the layer underneath it.