3 min read

Every AI word of 2026, month by month

19 AI engineering words on Hacker News, GitHub and AI Engineer talks since 2023, and which ones are still growing

Every few months the AI crowd gets a new word, prompt engineering, RAG, MCP, vibe coding, skills, and in 2026 there has been a new one every couple of weeks, harnesses, loops, personal agents, autoresearch, tokenmaxxing, software factories. I wanted to know which of these are still going, so for 19 of them I counted Hacker News stories and comments every month since January 2023, new GitHub repos with 10+ stars that use the word in their name or description, and talks at this year’s two AI Engineer conferences, World’s Fair in July and Code Summit this week.

Hacker News stories and comments per month. Each row is scaled to its own busiest month.

Table
TermBusiest monthMentions thenLast month
Prompt engineeringApr 202331544
RAGMay 2024508269
Context engineeringJul 202515016
Agent loopsJan 202616069
Vibe codingFeb 20261,7721,213
SkillsFeb 202628199
Personal agentsFeb 20262,963450
Agent managersFeb 202613551
MCPMar 20261,9871,344
GraphsMar 20269446
AutoresearchMar 202615718
WorktreesMar 2026222128
Agent memoryApr 20265934
TokenmaxxingJun 202614225
SubagentsJul 2026548437
HarnessAug 2026250226
SandboxesSep 2026258258
EvalsOct 2026241238
Software factoryOct 2026190117

Pretty much every 2026 word peaked by April. Personal agents (OpenClaw, Grok Bot, Meta’s Muse, ChatGPT’s dot, Hermes Agent) went from 2,963 Hacker News mentions in February to 400 to 450 a month since August, and the mix changed a lot underneath, OpenClaw went from 2,747 mentions to 143 while Grok Bot went from none to 128, and Muse and dot showed up in October. New personal agent repos went from 938 in February to 126 in October, which looks like people stopped building their own once the labs shipped one. Autoresearch peaked in March when Karpathy put out his repo, agent loops in January, agent managers in February and tokenmaxxing in June. The only words that peaked in the last few months are subagents (July), harness (August), sandboxes (September), evals and software factory (October).

Hacker News mentions and new GitHub repos with 10+ stars per month, each line scaled to its own peak.

Hacker NewsNew repos

Table
TermHN, last 3 monthsNew repos, last 3 months
Personal agents1,318434
Agent managers14649
Agent loops22438
Skills2641,141
Graphs13169
Harness673643
Sandboxes678207
Evals65596
Software factory40338

Talking about a word and building with it don’t happen at the same time. Repos that call themselves a harness went from 73 in January to 311 in April, right after Mitchell Hashimoto and OpenAI started using the term in February, and Hacker News caught up later with 250 mentions in August. DeepSeek Harness came out in August and got about 1,100 plugin repos that month alone, so I left those out of the harness count. Software factory goes the other way, 190 Hacker News mentions in October, four times what it had in May, and only 14 new repos. Tons of people are writing about factories and very few are open sourcing one, the ones that exist are things like fabro, Fusion and vercel-labs/eve-software-factory-template.

Share of AI Engineer talks whose title, abstract or track mentions the word.

World's Fair, Jul 2026Code Summit, Nov 2026

Table
TermWorld's Fair, Jul 2026Code Summit, Nov 2026
Verification14%21%
Coding agents8%24%
Agent teams7%14%
Software factory7%11%
Harness10%12%
Sandboxes6%8%
Skills10%10%
Memory11%8%
Evals9%2%
MCP7%4%
Context engineering3%0%
Personal agents7%2%
Autoresearch4%0%
RAG3%0%

The conferences moved the same way. At World’s Fair in July, 14% of talks mentioned verification or trust in the title, abstract or track, at Code Summit 21% do, agent teams went from 7% to 14% and software factories from 7% to 11%, while MCP went from 7% to 4% personal agents from 7% to 2% and autoresearch from 4% to none. Code Summit is a coding conference so coding agents jumping from 8% to 24% is not a surprise, and evals is the weird one, down from 9% to 2% at the conference while Hacker News talks about evals more than ever. The rest of the list reads like the Hacker News chart, people went from getting one agent to work to running a bunch of them without watching, with talks like “Building a Safe Software Factory”, “12-Factor-Factories”, “Code review without reviewing code” and “Sandboxes are dead”.

Last three months over the three before. Right of 1× is growing.

Hacker NewsNew repos

Table
TermHacker NewsNew repos
Software factory3.0×1.8×
Sandboxes1.6×0.8×
Subagents1.0×1.1×
Harness1.0×0.9×
MCP1.2×0.8×
World models0.9×0.9×
Personal agents1.1×0.7×
Evals1.2×0.6×
Worktrees0.9×0.8×
Skills0.9×0.8×
Vibe coding1.0×0.7×
RAG1.1×0.6×
Agent memory1.1×0.4×
Agent loops0.8×0.5×
Autoresearch0.9×0.4×
Tokenmaxxing0.3×–
Graphs0.7×0.5×
Agent managers0.8×0.4×

Putting the last three months next to the three before, software factory is the only word clearly growing on both, 3x the Hacker News mentions and 1.8x the repos. Sandboxes are up 1.6x on Hacker News. Pretty much everything else is flat or falling, and personal agent repos are at 0.7x.

Share of each month's new AI repos (300+ stars) that make video, animation or motion graphics.

  1. hypit-ai/hypit 28,656 stars

    Clone any viral video with AI agents. Not just a script, the whole workflow: swap the face, the words, the B-roll, ship 100 variants in one

  2. img2threejs/img2threejs 20,975 stars

    Rebuild the object in a reference image as a code-only, procedural, quality-gated, animation-ready Three.js model. Token-efficient image-to-

  3. Vincentwei1021/video-shotcraft 17,960 stars

    AI video skill for Claude Code & Codex — cinematic, editable product promos with Remotion: 124 shot recipe cards, 150 motion previews, beat

  4. yihui-dev/awesome-opus5-5-videos 14,406 stars

    A growing collection of viral videos made with Claude Opus 5.5 and the prompts behind them. Watch each original next to a live remake on Ski

  5. lemomo-ai/lemo-opuscar 13,748 stars

    A toolbox of film styles made in code for Claude Code: 66 ready to use, and your own. Each style is a style prompt plus a demo film made ent

Table
MonthVideo reposNew AI reposShare
Jan 202541323%
Feb 202571664%
Mar 202522481%
Apr 202581944%
May 202561753%
Jun 202561783%
Jul 2025111418%
Aug 202551424%
Sep 202511301%
Oct 202571834%
Nov 202551543%
Dec 202591925%
Jan 2026163355%
Feb 2026204315%
Mar 2026255255%
Apr 2026173735%
May 2026152915%
Jun 2026212449%
Jul 20262721113%
Aug 2026182467%
Sep 20262317513%
Oct 20262318412%

The one thing in the repos that Hacker News barely talks about is video. I went through the descriptions of the 8,088 AI repos with 300+ stars from the last post, and the share of new ones that make videos, animations or motion graphics went from 5% for the first five months of 2026 to 13% and 12% in September and October. Most of them are skills that get Claude Code or Codex to make videos in code with Remotion, three.js or ffmpeg. Hacker News mentions Remotion about 10 times a month and Bluesky about 150, and I keep seeing these videos on Instagram and X all the time.

If you want to learn more about your starred repos, you can run npx github-stars-organizer.

Hacker News counts are stories and comments from the Algolia API that contain the exact phrase. GitHub counts are repos created that month with 10+ stars and the word in their name or description, so the last couple of months are a bit low since new repos haven’t had time to get 10 stars. Personal agents is OpenClaw (and its old names clawdbot and moltbot), Grok Bot, Muse, ChatGPT dot, Hermes Agent and the phrases personal agent and autonomous agent. Counts as of November 8.