Every year, software reinvents its vocabulary. Most of last year's new words were about what AI could do. This year's words are about getting AI to work for business.
Here are the 10 words you couldn't escape in 2026, translated into plain English, with a light reality check.
1. Agentic Engineering
Basically what vibe coding is called once someone is accountable for it.
Agents write the code, while humans drive the architecture, review, and decide what gets shipped. Simon Willison spent the year pushing the term, and Google's Logan Kilpatrick drew the line between the two: "AI Studio is for vibe coding, Antigravity is for agentic engineering."
2. Harness Engineering
The setup around the model.
The harness is everything that turns a model into a working agent, including the tools it has access to, permissions on what it can do, memory, tests, and logs you can read when something goes wrong. OpenAI wrote about the idea in February, and by the summer it had its own track at the AI Engineer World's Fair.
While you rent the model from a provider, you can either build your own harness, or you could fork an existing one from GitHub and customise it to your needs.
3. Forward Deployed Engineer (FDE)
An engineer who works from inside your business.
Palantir created the role to send their engineers to customers, to work in their office, their systems and their data, and stay there until things run in production. In 2026, OpenAI and Anthropic were hiring them too, and people are calling it "the hottest role in AI".
4. Evals
Tests for whether the AI (model/agent/tool/harness) is any good.
An eval is a set of example tasks with known good answers, run every time you change the prompt, the model or the tools. The score tells you whether a change helped or made things worse.
Ask a vendor to tell you about their evals. If they can't, they probably don't have any, and can't measure whether their product will get better or worse next month.
5. Agent Skills
Repeatable instructions written for an agent to pick up when it needs them.
A skill is a folder with a SKILL.md file, plus any scripts or templates, that teaches an agent one job. Sort of like an SOP written for humans, except that it contains code and examples of good and bad output. Anthropic released the format in late 2025, and by 2026 Codex, Copilot and Gemini CLI could use it too.
If your team pastes the same process notes into a chat every week, write them up once as a skill your business owns. You don't even need to write it yourself, just use AI to create the skill for you.
6. Spec-Driven Development
Agreeing what to build before the agents start.
You write a specification, get it reviewed, and then agents build against it. Tools such as GitHub's Spec Kit and Amazon's Kiro were made for this.
Writing code takes minutes now, so the spec is where we spend way more time. Anyone who starts building after one call and a loose prompt will hand you a sloppy product.
7. Workslop (and the Slop Grenade)
Last year's slop, now in your inbox.
Workslop is AI-generated output that *looks* finished, but leaves the reader to do the thinking, like a report with no conclusion. Researchers writing in Harvard Business Review named it in 2025.
Taking this further, a "slop grenade" is unchecked AI output sent to a colleague for them to deal with. Harry Brundage coined it, and Shopify's Tobi Lütke credited him in public and asked people to make it a thing.
When someone says AI made them faster, check who is doing the rework. The hours sometimes land on someone else's desk, perhaps even yours?
8. Meat Proxy
The person standing between you and the chatbot.
A meat proxy forwards AI output without reading, understanding or checking it, so all they add is an extra hop between the model and you. Niklas Gruhn's August essay, "Don't be a meat proxy" made it go viral.
It sits next to workslop, where workslop is the unread work, and the meat proxy is the person who sent it on.
If a supplier's update reads as if it went straight from a chatbot to your inbox, ask them to talk you through it on a call. You're paying for their judgement, and you can ask the chatbot yourself for free.
9. AI-Native
The adjective on every 2026 pitch deck.
An "AI-native" product promises AI actually *doing* the work, as opposed to an existing app with a chatbot duct-taped to it.
Imagine the product with the AI removed. If it still works fine, it was never AI-native.
10. Coding Agents
Last year's copilots suggested the next line of code, where today's coding agents like Claude Code, Codex or Cursor perform a whole task. A coding agent takes your instructions, edits files, runs commands, reads errors, and keeps going until it either gets stuck or gets the job done. Take it a step further by adding tools, skills, evals and you've got yourself a harness.
Your build partner is probably using them already. Ask how they QA the agent work before it reaches you.
Honourable Mentions
- Context Rot: when answers get worse as you stuff more into the model's context.
- Ralph Loop: an agent run in a loop on one goal until it gets the job done.
- GEO / AEO: getting cited in AI answers as well as ranked on Google. It's still just SEO by the way.
- Agent Washing: ordinary automation rebadged as an "agent". More often than not, it's just a script with nothing "ai" or "agentic" about it.
- Vibe Engineering: Simon Willison's rival name for agentic engineering.
- Clanker: a meme slur for robots and chatbots, borrowed from Star Wars battle droids, which broke out in 2025 and stuck around.
- Tokenmaxxing: burning as many AI tokens as possible to look busy, or because you pay for Claude Ultra and you like getting your money's worth.
- AI;DR: "AI; didn't read", the slop era's TL;DR younger sibling to get out of reading anything that looks like AI slop.
- WebMCP: a proposed browser standard that lets a website tell an agent what it can do, so the agent stops guessing which button to click
- Service-as-Software: paying a supplier for the outcome, such as support tickets resolved, instead of for software your team has to run
The Big Takeaway
2025's buzzwords were mostly about what models could do. 2026's are about making them reliable (harnesses, evals, specs, skills) and about people who stay until it works, which is what FDEs and service-as-software both sell. Watch out for workslop, meat proxies and slop grenades though.
At Celestify, that work is the job. If you want the agents and someone accountable for them, let's talk.