What's hyped now, what already peaked, and which company you're in

The wave of decision models started with TypeSafe's Jev on 15 September 2026, and eight weeks later we list 25 of them, with OpenAI's on 6 October, Microsoft's on 9 October and Google's on 11 November ^[1]. The name only dates from September (I wrote about why everyone wants your decisions in October), so a kind of product went from a name to 25 launches in eight weeks, and that's the speed I'm holding everything else here against.

I started with a list of 24 things, 12 that I thought were hyped and 12 that I thought had already peaked, and when I went looking for something to back each one, 10 didn't hold up, so they're not here. The first version of this post, which went up earlier on 13 November, also said that 25 of our 27 decision models came out in the last six months, which was true and silly when the name is two months old, and it had a chart built on the same kind of sum, and both are gone. What's left is sorted by who's talking about it, because a frontier lab or an AI-native team, a company that went all in on AI this year and a tech company that's behind are having three different conversations. The sorting is mine and I have no survey behind it.

What was being argued about in May

On 28 May Fortune called tokenmaxxing over, after a spring in which big companies ranked engineers by tokens spent, Meta's internal leaderboard was reported in April and came down two days later, and Uber got through its whole 2026 token budget in four months ^[2]. Two days earlier Kilo Code had written up what its own engineers do, after seeing posts on X and LinkedIn from people who said they ran 50 to 100 coding agents at once ^[3]. And at the start of the month MCP.Directory was explaining how to stop MCP servers eating the context window, a problem someone had measured in November 2025, when seven servers in Claude Code took 67,300 tokens of tool definitions before anyone typed a word, a third of a 200,000-token window ^[4]. Tokens, agents and tools are all counts, and I wrote on 7 November that it's never been about quantity.

The frontier and the AI-native teams

Start with harness engineering. A preprint by Chaoyue He and colleagues, first posted in March 2026, credits OpenAI with the label, for long-running systems whose reliability depends on repository maps, AGENTS.md files, architectural rules and runtime controls more than on how the prompt is worded ^[5], and the study I wrote about on 7 November found that changing the harness moved coding scores about 7.8 times as much as changing the model ^[2]. I think that's why the argument about which model is best matters less than it did, since the model is one setting among several.

The cost they count is the cost of a finished task. DoiT published a way of measuring it on 13 August 2026, on the grounds that a cheaper price per token can mean a dearer task when the model takes more steps or fails more often ^[6] (DoiT sells the tooling for it, and says its worked examples are illustrative).

Coding agents run in parallel, and fewer of them than the posts say. Of the seven senior engineers Kilo asked, most run two to four in the foreground with a handful more in the background, and Kilo's conclusion was that the 100-agent posts are mostly there for attention ^[3]. We run about 50 research agents and one coordinator for our own listings ^[7], so I'm not saying a big team of agents can't work, and I wrote about Anthropic's thousand, which Anthropic itself called a toy experiment ^[8].

The newest thing they talk about is weight-space programming. On 2 November four researchers at MIT CSAIL, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, wrote up GPT-6 Astra reading the weights of a small Qwen model and editing 0.016% of them, with no training, which took its MATH-500 score from 7.9% to 48.1% ^[9]. They call the results very early, and on a second model the edit couldn't add an ability the model lacked ^[9]. I wrote about it the next day, and I still think it's a good step and a long way from the singularity.

Open-weight models are on this side too. Of the ones we list, 11 first shipped in July, 16 in August, 9 in September and 22 in October ^[10]. That's a floor, since it only counts the ones we list, and we only started listing them on 20 October, so I wouldn't read a trend into the months.

The companies that embraced AI

Forward deployed engineers are what these companies are hiring. TechCrunch reported on 30 July 2026 on a study by the search firm Christian & Timbers, in which the share of companies planning to hire them went from between 5 and 10% at the start of the year to 70% by the end of June ^[11].

They also have a review problem, because the writing worked. In Sonar's survey of more than 1,100 enterprise developers, as Built In reports it, 42% of committed code is AI-generated or AI-assisted, up from 6% in 2023, and more than a third say it takes more effort to review than code a person wrote ^[12]. In a Flux survey of 309 engineering leaders and practitioners that LeadDev covered on 30 June, 35% of teams write code with AI and won't ship it, because they can't judge the risk ^[13]. That's the gap AI code review is sold into.

Chief-of-staff agents are the other thing they're buying. Asana launched Dash on 4 June at its Work Innovation Summit in London and calls it a personal AI chief of staff, which picks follow-ups out of meetings, Slack threads and email and turns them into tracked work ^[14]. Desktop agents, with a person confirming the risky steps, went from demo to product in the same months, and Google made its Computer Use tool generally available on the Gemini API on 30 October ^[15].

The companies that are behind

If the conversation at your company is one of these, I think it's behind. Fast Company wrote on 6 May 2025 that prompt engineering as a job had all but disappeared, two years after it was the hot new one, and an economist at Indeed told them it was never a big enough job title to track ^[16]. Which model is best is on this side for the reason harnesses are on the other. Token leaderboards, and bragging about how much code your agents wrote, went with tokenmaxxing in May, and Microsoft told its engineers in August that tokenmaxxing isn't what it's optimising for ^[2].

Connecting every MCP server you can find went earlier than that. The fixes came in November 2025 and February 2026, Anthropic's tool search, which by Anthropic's figure cuts token use by 85%, and Cloudflare's Code Mode, which Cloudflare says takes 99.9% off the input tokens for its own API, as MCP.Directory reports both ^[4] (I don't think MCP is the future anyway, I think plain APIs win). On our listing for Railway's MCP server, loading only the two tools a log-reading job needs costs about 1,900 tokens against about 52,000 for the whole list ^[17].

How fast it's moving

Prompt engineering took about two years to go from the hot new job to all but gone ^[16], the token leaderboards were news in April and called over by the end of May ^[2], and decision models are eight weeks old with 25 on our list ^[1]. I think the gap between a thing getting its name and everyone having one is getting shorter, though three examples isn't a measurement.

The two from before September are an IBM model from May and a classifier from 2024 that do the same job and sit in the same category ^[1].

What's missing here is our own numbers for the rest. About half of the agent harnesses, agent frameworks and code sandboxes we list have no first release date on their listing yet ^[18], so I can't tell you from our data whether frameworks have peaked or sandboxes are rising, and both came off the list until I can.

If you want to know which company you're in, write down the last three things your team argued about on AI and see which section they fall in.

References

Read on 13 November 2026. Quotes are as published.

1. [1]Anchor Terminal, decision model listings, each with its first release date and the vendor pages it was graded from, counted on 13 November 2026. Source for Jev on 15 September, the 25 that shipped from then to 11 November, the dates for OpenAI, Microsoft and Google, the count by month, and the two older models in the category. anchorterminal.com
2. [2]Anchor Terminal, A machine that runs all night, and why the labs aren't at the forefront, 7 November 2026, with Fortune's reports of 9 April and 28 May 2026, LeadDev's of 17 August 2026 and Zhang and others (arXiv 2605.23950) as its sources. Source for Fortune calling tokenmaxxing over, the token leaderboards, Meta's coming down two days after it was reported in April, Uber's budget, Microsoft's message to engineers, and the harness moving coding scores about 7.8 times as much as the model. anchorterminal.com
3. [3]Kilo Code blog, Darko Gjorgjievski, 26 May 2026. Are parallel coding agents actually useful, or just Twitter theater? Source for the posts claiming 50 to 100 agents, the seven senior engineers asked, the two to four foreground agents most of them run, and Kilo's conclusion. Kilo's own account of its own team. blog.kilo.ai
4. [4]MCP.Directory, 4 May 2026. MCP context bloat, the 2026 fixes that work. Source for the measurement filed on 10 November 2025 (seven MCP servers, 67,300 tokens, 33.7% of a 200,000-token window), Anthropic's tool search of 24 November 2025 and its 85% figure, and Cloudflare's Code Mode of 20 February 2026 and its 99.9% figure for the Cloudflare API, as reported there. mcp.directory
5. [5]He, C., Zhou, X., Wang, D., Xu, H., Liu, W. and Miao, C., first posted March 2026, version 2 posted 23 April 2026. Harness engineering for language agents, a preprint that isn't peer reviewed. Source for OpenAI's use of the label harness engineering and what it covers. preprints.org
6. [6]DoiT, 13 August 2026. Cost per task, not cost per token. Source for the framework, the argument that a cheaper token can mean a costlier task, and its case studies being illustrative. doit.com
7. [7]Anchor Terminal, how our listing research runs, November 2026, not published. Source for about 50 research agents and one coordinator.
8. [8]Anthropic, claude.dev blog, 9 November 2026. Safely scaling multi-agent teams with Claude Fable 5.2. Source for the 1,000-agent run being described as a toy experiment. claude.dev
9. [9]Takehi, R., Teoh, J., Zhang, A. and Khattab, O., MIT CSAIL, 2 November 2026. Sparks of weight-space programming. Source for GPT-6 Astra editing Qwen3.5-0.8B, the 0.016% of weights, MATH-500 going from 7.9% to 48.1%, the authors calling the results very early, and the edit not adding an ability to Gemma 3 270M. rikiyat.github.io
10. [10]Anchor Terminal, open-weight model listings, each with its first release date, counted on 13 November 2026, the first of them listed on 20 October 2026. Source for the count of listed models that first shipped in July, August, September and October 2026. anchorterminal.com
11. [11]TechCrunch, Rebecca Bellan, 30 July 2026. Forward-deployed engineers are the AI industry's latest talent obsession. Source for the Christian & Timbers study as TechCrunch reports it, with the share of companies planning to hire forward deployed engineers going from 5 to 10% to 70%. techcrunch.com
12. [12]Built In, Lauren Hanford, 2026. 96 percent of developers don't trust AI-generated code. Source for the 2026 State of Code developer survey of more than 1,100 enterprise developers, the 42% and 6% figures and the review effort, as reported there. builtin.com
13. [13]LeadDev, Chantal Kapani, 30 June 2026. AI-generated code sparks production confidence crisis. Source for the Flux survey of 309 engineering leaders and practitioners and the 35% of teams who use AI to write code and won't ship it, as LeadDev reports it. leaddev.com
14. [14]The New Stack, 4 June 2026. Asana says its new AI chief of staff turns your Slack chaos into trackable work. Source for Dash, where it was launched, Asana's description of it and what it does. thenewstack.io
15. [15]Google, Gemini API changelog. Source for the Computer Use tool reaching general availability on 30 October 2026. ai.google.dev
16. [16]Fast Company, Henry Chandonnet, 6 May 2025. On prompt engineering quickly going extinct. Source for prompt engineering as a job having all but disappeared two years after it was hailed as a hot new one, and Allison Shrivastava of the Indeed Hiring Lab on the title never being large enough to track. fastcompany.com
17. [17]Anchor Terminal, the Railway MCP server listing. Source for the token counts of the whole tool list and of the two tools a log-reading job needs. anchorterminal.com
18. [18]Anchor Terminal listings data, counted on 13 November 2026, not published as a table. Source for 25 of 56 agent harnesses, 10 of 20 agent frameworks and 10 of 22 code sandboxes having no first release date.