Milestones
55 dated events that mattered but were not model releases, 2017–2026
Not everything that shaped this history was a set of weights. A milestone records a dated event — a product launch, a statutory regime — that belongs on the timeline but has no parameters, context window or licence. Every milestone needs a primary source, exactly like a model record.
This is a deliberately short list, and it is short because the bar is high: an event earns a record here only when it changed what the labs in this dataset could ship or sell. 55 records is the honest count, not a placeholder for a longer one.
Significance
Kind
Recorded against
No milestone matches those filters.
2017 1 milestone
-
Google Brain published the Transformer architecture, replacing recurrent neural networks with self-attention mechanisms. The paper demonstrated that attention alone, without recurrence or convolutions, could achieve state-of-the-art results on machine translation.
Every large language model since — GPT, BERT, PaLM, Gemini, Llama — is a Transformer. The architecture is the foundation of the entire field. Without it, there is no modern AI.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 95/100
2019 1 milestone
-
OpenAI demonstrated that scaling to 1.5B parameters produced qualitatively better text generation, then withheld the full model citing concerns about malicious applications — the first staged "too dangerous to release" decision by a major AI lab.
First public debate about AI safety and responsible publication by a major lab. Also demonstrated that scaling produces emergent qualitative improvements, not just quantitative ones.
- Model line
- GPT · 18 tracked releases
- Most recent tracked release in that line before this date
- None. This dataset's record of the GPT line starts later, with GPT-4 on .
Record status: verified · confidence 90/100
- openai.com Official announcement archived
2020 2 milestones
-
OpenAI researchers showed that language model loss follows predictable power laws across parameters, dataset size, and compute — spanning more than seven orders of magnitude.
Gave the field a quantitative framework for predicting how much compute a model needs. Directly predicted GPT-3's optimal size four months later. Compute allocation became a strategic decision rather than guesswork.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 95/100
-
OpenAI trained GPT-3, a 175B-parameter autoregressive language model that demonstrated in-context few-shot learning without fine-tuning — sometimes becoming competitive with prior state-of-the-art fine-tuning approaches.
Proved that massive scale enables a qualitatively different capability: few-shot learning. Marked the transition from research curiosity to commercial viability. The foundation-model paradigm starts here.
- Model line
- GPT · 18 tracked releases
- Most recent tracked release in that line before this date
- None. This dataset's record of the GPT line starts later, with GPT-4 on .
Record status: verified · confidence 95/100
2022 4 milestones
-
OpenAI fine-tuned GPT-3 with human feedback (RLHF) to follow instructions. The 1.3B-parameter InstructGPT model was preferred by human evaluators over the 175B GPT-3.
The alignment tax became the alignment dividend. Fine-tuning with human feedback made language models actually useful — not just capable at next-token prediction, but capable at doing what humans want. Every chat product after this uses this technique.
- Model line
- GPT · 18 tracked releases
- Most recent tracked release in that line before this date
- None. This dataset's record of the GPT line starts later, with GPT-4 on .
Record status: verified · confidence 95/100
-
DeepMind showed most existing LLMs (including GPT-3) were significantly undertrained — too many parameters, too little data. For compute-optimal training, model size and training tokens should scale equally.
Corrected the Kaplan scaling laws. Reshaped how the entire field allocated compute budgets. Directly inspired LLaMA, Mistral, and the "smaller but better trained" open-source ecosystem.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 95/100
-
Harrison Chase published the first version of LangChain as a small Python package for chaining LLM calls, prompt templates and external tools.
The first widely adopted framework for wiring a language model to tools, memory and multi-step control flow. Almost every 2023 agent experiment, including AutoGPT and BabyAGI, was built on or against it.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 70/100
- x.com Official announcement
- langchain.com Official announcement archived
- en.wikipedia.org Independent analysis archived
-
The public chatbot built on GPT-3.5 that brought large language models to the mainstream.
The moment large language models reached the general public. Everything in this timeline after November 2022 happens in a market ChatGPT created.
- Model line
- GPT · 18 tracked releases
- Most recent tracked release in that line before this date
- None. This dataset's record of the GPT line starts later, with GPT-4 on .
Record status: verified · confidence 90/100
- openai.com Official announcement archived
2023 12 milestones
-
Meta released LLaMA (7B–65B parameters), trained exclusively on publicly available data. The weights were initially gated behind a researcher application form and leaked on March 3, 2023, igniting the open-weights fine-tuning ecosystem.
Proved that models trained on public data could match proprietary giants. The accidental leak created the open-weights fine-tuning ecosystem — Alpaca, Vicuna, Mistral, and the entire open-source LLM movement descend from this release.
- Model line
- Llama · 8 tracked releases
- Most recent tracked release in that line before this date
- Llama 1 —
Record status: verified · confidence 90/100
- ai.meta.com Official announcement archived
- arxiv.org Official documentation archived
-
Within two weeks, Stanford released Alpaca (LLaMA 7B fine-tuned on 52K synthetic instructions for $600) and LMSYS released Vicuna (LLaMA 13B fine-tuned on 70K ShareGPT conversations for $300). Both demonstrated that instruction-tuning a base model was trivially cheap and replicable.
Proved that instruction-tuning was trivially cheap. The "base model + synthetic instructions" recipe became the template for every open instruction-tuned model. Also introduced GPT-4-as-judge evaluation.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 70/100
- crfm.stanford.edu Official announcement archived
- lmsys.org Official announcement archived
- en.wikipedia.org Independent analysis archived
-
OpenAI released GPT-4, the first major model to accept images as input alongside text, demonstrating human-level performance on professional benchmarks including the bar exam.
Established the new performance ceiling for commercial LLMs. Introduced multimodal input. Every frontier model after GPT-4 was expected to be multimodal. The benchmark arms race intensified.
- Model line
- GPT · 18 tracked releases
- Most recent tracked release in that line before this date
- GPT-4 —
Record status: partly verified · confidence 85/100
- openai.com Official announcement archived
- theverge.com News reporting archived
-
Google's ChatGPT rival chatbot that brought the company into the consumer AI race.
Google's first consumer answer to ChatGPT, and the start of the assistant race between the two largest labs.
- Model line
- Gemini · 15 tracked releases
- Most recent tracked release in that line before this date
- None. This dataset's record of the Gemini line starts later, with Gemini 1.0 Ultra, Gemini 1.0 Pro and Gemini 1.0 Nano on .
Record status: verified · confidence 90/100
- blog.google Official announcement archived
-
OpenAI gave ChatGPT the ability to call third-party services, plus its own browsing and code-interpreter plugins, initially to a waitlist.
The first time a mass-market language model could act outside its own text box. It is the origin of the tool-use pattern every later agent is built on, and OpenAI said so directly: it expected open standards to emerge for exposing an AI-facing interface.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 90/100
- openai.com Official announcement archived
-
A roughly 100-line Python loop that creates, prioritises and executes its own tasks against an objective. Shared first as the 'Task-Driven Autonomous Agent', then released as the babyagi repository days later.
The minimum viable autonomous agent, small enough to read in one sitting. It defined the plan-execute-reprioritise loop that the whole 2023 agent wave copied, and its author explicitly framed it as documenting a pattern rather than shipping a product.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 80/100
- github.com Code repository archived
- yoheinakajima.com Official announcement archived
-
SG
An open-source program that gives GPT-4 a goal in natural language and lets it decompose the goal into sub-tasks, then run them against a browser, a file system and a code interpreter without further prompting.
The moment 'autonomous agent' became a mainstream idea rather than a research term. It became one of the fastest-starred repositories in GitHub history and set the expectations — and the disappointment — that the next two years of agent work were measured against.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 60/100
- en.wikipedia.org Independent analysis archived
- github.com Code repository archived
-
Baidu Intelligent Cloud unveiled Comate, a code assistant built on the Wenxin (ERNIE) model, at a Wenxin large-model technical conference in Chengdu and opened it for invited testing. It covered around 30 languages, and Baidu claimed roughly a 50% acceptance rate for its suggestions in internal use. Baidu gave it the Chinese name 文心快码 (Wenxin Kuaima) on 2024-06-28.
The first coding assistant shipped by a Chinese hyperscaler on its own foundation model, three weeks after GitHub Copilot Chat's own beta and four months before Alibaba's equivalent. It marks the start of the Chinese Copilot generation — completion and chat inside an editor, not an agent — which is the baseline everything in the 2025 Chinese harness wave was built against.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 70/100
- news.qq.com News reporting
- it.ithome.com News reporting
- comate.baidu.com Official documentation archived
- baike.baidu.com Independent analysis
-
OpenAI shipped gpt-4-0613 and gpt-3.5-turbo-0613, fine-tuned to decide when a described function should be called and to emit a JSON object matching its signature.
Tool use stopped being a prompting trick and became an API primitive with a defined contract. Every agent framework after this date builds on structured tool calls rather than on parsing free text.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 90/100
- openai.com Official announcement archived
-
Meta released Llama 2 with explicit commercial-use rights, co-launched with Microsoft as preferred partner and distributed through Azure. The first major LLM released for free commercial use.
First major LLM released with explicit commercial-use rights. Legitimised open-weight models for enterprise and product use. The open-weights ecosystem moved from research to production.
- Model line
- Llama · 8 tracked releases
- Most recent tracked release in that line before this date
- Llama 2 —
Record status: verified · confidence 90/100
- about.fb.com Official announcement archived
-
iFlytek announced iFlyCode 1.0 (星火飞码) at the same press conference as its Spark cognitive model V2.0, an in-editor assistant doing code generation, completion, explanation, error correction and unit-test generation. iFlytek reported that more than 2,000 of its own engineers were already using it.
The second Chinese coding assistant to ship, and the one that came bundled with a model launch rather than as a cloud product — iFlytek treated the assistant as the demonstration of Spark V2.0's coding ability. It is also the clearest case in this batch of a lab that never followed the assistant with an agent or a CLI, which is why iFlytek is absent from the 2025-26 harness wave.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 70/100
- ithome.com News reporting archived
- news.zol.com.cn News reporting
- xfyun.cn Official documentation archived
-
Alibaba Cloud unveiled 通义灵码 (Tongyi Lingma), a coding assistant built on its Tongyi code model, at the Apsara Conference in Hangzhou, demonstrating it generating a hundred-odd lines of Python from a one-line request. It supported Java, Python and Go with generation, chat, unit-test generation and code explanation, and the individual tier was free.
Alibaba's entry into the Chinese Copilot generation, and the direct ancestor of everything Alibaba later shipped in this space: the Lingma team became Qoder, and Lingma IDE now carries a Qoder CN banner. It is the assistant a Chinese lab kept iterating on longest before replacing it with an agentic product.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 65/100
- developer.aliyun.com Official announcement archived
- geekpark.net News reporting
- help.aliyun.com Official documentation archived
2024 9 milestones
-
Google retired the Bard brand, renamed the assistant Gemini, and shipped it as a mobile app with a paid Advanced tier.
The point Google stopped positioning its assistant as an experiment and started selling it. Bard had been the reactive answer to ChatGPT; Gemini was the product line it became.
- Model line
- Gemini · 15 tracked releases
- Most recent tracked release in that line before this date
- Gemini 1.0 Ultra, Gemini 1.0 Pro and Gemini 1.0 Nano —
Record status: verified · confidence 90/100
- blog.google Official announcement archived
-
Cognition unveiled Devin, an agent with its own shell, editor and browser, marketed as a teammate that completes engineering tasks end to end rather than as an autocomplete.
The first time an agent was positioned as a colleague with a job description rather than a developer tool. It reset what a coding agent was expected to do, and its SWE-bench claims made benchmark performance on real repositories the field's scoreboard.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 75/100
- cognition.com Official announcement archived
- siliconangle.com News reporting archived
-
The Qwen team published v0.0.1 of Qwen-Agent, its open framework for building applications on Qwen's function-calling, planning and memory: an agent loop, a tool registry, a code interpreter, RAG and a BrowserQwen Chrome extension, packaged to PyPI the same day.
The first agent framework published by a Chinese frontier lab under its own model's name — LangChain's role, but written by the lab that makes the model rather than around it. It is the line from which Alibaba's later first-party harnesses descend, and it predates Qwen Code by fifteen months.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 75/100
- github.com Code repository archived
- pypi.org Code repository archived
- qwenlm.github.io Official documentation archived
-
Regulation (EU) 2024/1689 — the first broad statutory regime for AI systems, including obligations that fall on general-purpose model providers.
The first time training and releasing a general-purpose model carried statutory duties rather than voluntary commitments. Obligations phase in over years, but this is the date the regime began to exist.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 85/100
- eur-lex.europa.eu Official documentation archived
-
Replit opened early access to an agent that takes a natural-language description and sets up the environment, writes the code, provisions a database and deploys the result.
The first agent that shipped a running application rather than a diff, and the first commercial evidence that agents could carry a business: Replit's revenue growth after this release is what pulled the rest of the industry into agent products.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 55/100
- x.com Official announcement archived
- aikatana.com News reporting archived
- replit.com Official announcement archived
-
OpenAI released o1, the first reasoning model, demonstrating that chain-of-thought inference at test time could dramatically boost performance on math, science, and coding benchmarks — scoring 83% on IMO problems versus GPT-4o's 13%.
Inference-time compute became a new scaling axis. Before o1: bigger model = better. After o1: smarter thinking = better. Every frontier lab subsequently developed reasoning models.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 90/100
- openai.com Official announcement archived
-
Anthropic put a public beta API in developers' hands that lets Claude look at a screen, move a cursor, click and type — operating software built for people rather than calling an API.
The first frontier model offered as a general computer operator. It opened the second track of the agent era: instead of every tool needing an API, the agent uses the interface a human would.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 90/100
- anthropic.com Official announcement archived
-
Zhipu AI launched AutoGLM, an agent that operates a phone or a web browser on the user's behalf from a spoken or typed instruction. Its published architecture splits a Planner from a Grounder so that deciding what to do and locating the pixel to tap are separate models — the paper reports 55.2% success on web browsing and 36.2% on Android control.
China's answer to computer use, launched three months before OpenAI's Operator and two days after Anthropic's computer-use API. It aimed at the phone rather than the desktop, which is the split that has defined Chinese agent products since: Western agents drive a browser, AutoGLM drove WeChat, Taobao and Meituan.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 65/100
-
Anthropic published MCP, an open standard and reference servers for connecting assistants to data sources and tools through one protocol instead of bespoke integrations.
The integration layer the agent era had been missing. Within a year it was adopted across competing labs and tools, making it the rare piece of AI infrastructure that became a shared standard rather than a moat.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 90/100
- anthropic.com Official announcement archived
2025 15 milestones
-
ByteDance shipped Trae, a VS Code fork with an AI builder mode, free to use and running on other labs' frontier models (Claude 3.5 Sonnet and GPT-4o) rather than ByteDance's own.
The first AI-native development environment from a major Chinese platform company, and the earliest Chinese entry in the coding-harness race. Like Manus a month later, it competed on the harness while renting the model, which is the pattern Chinese agent products followed until their own coding models caught up.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 50/100
- test-news.aibase.com News reporting
- visualstudiomagazine.com News reporting archived
- trae.ai Official documentation archived
-
DeepSeek released an open-source reasoning model matching OpenAI o1's benchmark performance, under MIT licence, at a fraction of the training cost. NVIDIA lost $589B in market cap on the following trading day.
First open-source reasoning model to match proprietary performance. Triggered a global reassessment of AI economics. Reasoning capability was no longer a proprietary moat.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 85/100
- deepseek.com Official announcement archived
- arxiv.org Official documentation archived
-
A research preview of an agent that drives a browser on the user's behalf — filling forms, ordering things, navigating sites — released to ChatGPT Pro subscribers.
The first consumer-facing agent from OpenAI that acted on the open web rather than answering questions. It moved computer-using agents from a developer capability to a subscription product.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 90/100
- openai.com Official announcement archived
-
An agent inside ChatGPT that runs multi-step web research over tens of minutes and returns a cited report rather than a chat answer.
Created the research-agent category. Within months Google, Anthropic, Perplexity and xAI had shipped their own, and 'deep research' became a generic product name rather than a brand.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 90/100
- openai.com Official announcement archived
-
Anthropic released Claude Code as a limited research preview: an agentic coding tool that works in the terminal, reading and editing a repository and running commands.
Established the terminal as the agent's home rather than the IDE, a shape that Google, OpenAI and xAI all copied within eighteen months. It also became the reference harness that later open-source agents were benchmarked against.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 85/100
- anthropic.com Official announcement archived
-
BE
A general-purpose agent from a Chinese startup, launched invitation-only, that plans and executes long multi-step tasks in a cloud sandbox and returns finished deliverables. It ran on other labs' models rather than its own.
Showed that a competitive agent product could be built on top of someone else's frontier model, and that the agent layer was a separate market from the model layer. Its viral launch is also the moment agents became a mass consumer expectation outside the US labs.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 65/100
- en.wikipedia.org Independent analysis archived
-
OpenAI reused the Codex name for something new: an agent that runs in its own cloud sandbox, works several tasks in parallel against a repository, and returns pull requests.
Moved coding agents off the developer's machine. Work being dispatched to a fleet of remote agents rather than supervised in an editor is the shape that OpenAI, Google and GitHub all converged on through 2025.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 85/100
- openai.com Official announcement archived
-
At Baidu AI Open Day, vice-president Chen Yang launched Comate AI IDE, a standalone AI-native development environment for 文心快码 rather than an editor plug-in, with multi-agent collaboration, MCP support, one-click conversion of a design file into code and ERNIE 4.0 X1 Turbo built in. Baidu said Comate was by then writing over 43% of its own new code.
The moment Baidu stopped shipping a plug-in and shipped an environment — the same move Alibaba made two months later with Qoder and Tencent a month after that. Its distinguishing bet was multimodal input: design-file-to-code as a first-class path, which none of the CLI-shaped Chinese harnesses attempted.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 75/100
- qbitai.com News reporting archived
- stdaily.com News reporting
- developer.baidu.com Official documentation archived
- oschina.net News reporting archived
-
Google released an Apache-2.0 licensed agent that runs in the terminal against a local codebase, with a free tier for personal Google accounts.
The first terminal agent from a frontier lab released as open source with a generous free tier, which turned the agent harness from a paid product into a commodity distribution channel for the underlying model.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 90/100
- blog.google Official announcement archived
-
ByteDance's Trae team open-sourced Trae Agent under MIT: the LLM-based software-engineering agent behind its IDE, released as a research-oriented CLI with a pluggable model layer and a published SWE-bench harness.
A frontier-scale Chinese lab publishing its production coding agent as a research artifact rather than a product. It gave the open SWE-agent literature a lab-maintained baseline, and it is the earliest of the Chinese first-party harnesses to be released as source rather than as a download.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 65/100
- github.com Code repository archived
- aibase.com News reporting archived
- marktechpost.com News reporting archived
-
OpenAI merged Operator's browser control and deep research's multi-step reasoning into a single agentic mode inside ChatGPT, with its own virtual computer and terminal.
The point the separate agent experiments were folded back into the main product. After this, an agent was a mode of the assistant rather than a distinct thing you went somewhere else to use.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 85/100
- openai.com Official announcement archived
-
Alibaba open-sourced Qwen Code alongside Qwen3-Coder: an Apache-2.0 terminal coding agent adapted from Google's Gemini CLI, with the prompts and parser reworked for the Qwen3-Coder function-call format.
The first frontier-lab harness that was a fork of another frontier lab's harness. Google had released Gemini CLI open-source four weeks earlier; Alibaba took it, retargeted it at its own model and shipped it the same day as the model — which is the moment the open terminal agent became shared infrastructure rather than each lab's own product.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 80/100
- qwenlm.github.io Official announcement archived
- github.com Code repository archived
- alibabacloud.com Official announcement
-
Alibaba launched Qoder, an agentic coding platform, into free global public preview. It pairs a proprietary Next-Edit-Suggestion model with third-party frontier models — Claude, Gemini and GPT — behind a scheduler that picks one per task, and splits work between an Agent Mode for small edits and a Quest Mode that runs an end-to-end task autonomously.
The first Chinese lab to ship a coding agent aimed at developers outside China from day one, and the one that got the scale: reported at 100,000 users in five days and around five million by mid-2026. It is also the batch's clearest admission that the harness and the model are separable — Alibaba's own agentic IDE routes to Anthropic, Google and OpenAI models alongside its own.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 70/100
- accessnewswire.com Official announcement archived
- qoder.com Official documentation archived
- finance.yahoo.com News reporting archived
-
Tencent Cloud released CodeBuddy Code, a terminal coding agent driving the whole write-debug-test-deploy loop from natural language, and on the same day opened CodeBuddy IDE to public beta. The domestic build runs on DeepSeek with unlimited use; the international build routes to GPT and Gemini. IDE and CLI share one quota.
The day Tencent completed the set — plug-in, IDE and CLI — and the first Chinese vendor to hold all three shapes at once. It is the clearest example of a Chinese harness shipped without a matching first-party model in front of it: Tencent has Hunyuan, and pointed its own CLI at DeepSeek anyway.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 75/100
- brands.cnblogs.com Official announcement
- ithome.com News reporting archived
- news.qq.com News reporting
- codebuddy.ai Official documentation archived
-
Moonshot AI published Kimi CLI, an Apache-2.0 terminal agent written in Python that drives the Kimi models against a local codebase. It was renamed Kimi Code CLI in January 2026 and later rewritten in TypeScript.
Moonshot's entry into first-party terminal agents. The date is the earliest release Moonshot published, not the day the tool first existed — the tag is numbered 0.33, so earlier versions shipped untagged and the true first availability is earlier than this record can establish.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 45/100
- github.com Code repository
- moonshotai.github.io Official documentation archived
- platform.kimi.com Official documentation archived
2026 11 milestones
-
OC
Peter Steinberger's self-hosted personal agent — first published in late 2025 under earlier names, renamed Clawdbot then Moltbot then OpenClaw — that runs on the user's own machine and is driven from WhatsApp, Telegram, Discord, Slack or Teams.
The first viral consumer agent that was not a lab product: open source, self-hosted, and reached through chat apps people already had. It also made prompt injection and agent permissions a mainstream security story rather than a research concern.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 75/100
- openclaw.ai Official announcement archived
- en.wikipedia.org Independent analysis archived
-
NR
Nous Research released an open-source, model-agnostic agent that runs as a persistent daemon on the user's own infrastructure, keeps memory across sessions and writes reusable skill files from its own experience.
The self-improving open-source harness: it made the agent's accumulated memory and skills, rather than the model behind it, the thing that got better over time. It became the fastest-growing open-source agent project of 2026.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 60/100
- hermesatlas.com Independent analysis archived
- marktechpost.com News reporting archived
- hermes-agent.nousresearch.com Official documentation archived
-
xAI's first agentic coding tool: a terminal-native CLI that runs an autonomous agent against a local codebase, opened in early beta to top-tier subscribers before widening later in May.
The last of the frontier labs to ship a coding agent, which is the point at which the terminal harness became table stakes rather than a differentiator.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 55/100
- buildfastwithai.com Independent analysis archived
- techloy.com News reporting archived
- channelinsider.com News reporting
-
Moonshot AI republished Kimi Code CLI as a new MIT-licensed repository, rewritten from Python to TypeScript, installable without Node.js, with coder, explore and plan subagents that run in isolated contexts. It migrates the configuration and sessions of the older Kimi CLI it replaces.
The subagent-with-its-own-context pattern arriving in a Chinese first-party harness, and a relicensing from Apache-2.0 to MIT that removed the last friction from redistributing it. Moonshot's own framing — 'the starting point for next-gen agents' — positions the harness, not the model, as the product.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 55/100
- github.com Code repository archived
- marktechpost.com News reporting archived
- moonshotai.github.io Official documentation archived
-
MiniMax renamed its desktop agent app to MiniMax Code and added a `minimax` command-line shortcut, turning a general-purpose assistant into a coding harness. Its agent loop was rebuilt on the open-source OpenCode and Pi Agent projects and trained jointly with the MiniMax M3 model released three days later.
The clearest case of the harness being treated as part of the model rather than a wrapper around it: MiniMax states that Code was designed for M3 and trained together with M3. It is also the batch's one proprietary Chinese harness — a desktop download, not a repository — built on other people's open harnesses.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 70/100
- agent.minimaxi.com Official documentation archived
- minimaxi.com Official announcement archived
-
Z.ai released ZCode, a free desktop 'agentic development environment' built around GLM-5.2, on macOS, Windows and Linux, with bring-your-own-key support for third-party models and a 1.5× quota bonus for GLM Coding Plan subscribers.
The point a Chinese lab stopped competing on model price and started competing on the harness. Z.ai pitched ZCode directly against Cursor, Claude Code and Copilot and gave the app away to pull developers onto its own coding subscription — the same distribution play Google made with Gemini CLI, run by a lab whose model was the cheaper one.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 60/100
- scmp.com News reporting archived
- x.com Official announcement
- venturebeat.com News reporting archived
-
xAI launched Grok Bot in beta: always-on agents that share a cloud computer of their own, sign into the apps and websites a team already uses, remember past work and keep going without a prompt, returning when something needs approval. Available to SuperGrok and Cursor subscribers on desktop and iOS, with its own usage allowance.
The shift from an agent you prompt to one that persists: Grok Bot has a computer, a memory and work that starts without being asked. It is a product built on Grok models, not a model, so it is recorded as an agent milestone.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 80/100
-
DeepSeek open-sourced its first first-party agent harness under MIT as a v0.1 developer preview. Built on a dependency-injection core called Cordis, every layer — model adapter, tool registry, session log, sandbox, web UI and the agent loop itself — is a replaceable plugin, and it ships adapters for Anthropic, OpenAI, Bedrock, Azure and Google rather than only DeepSeek's own models.
The last of the large Chinese labs to ship a harness, and the one that refused to make it a product: released as a preview, provider-agnostic by design, with DeepSeek's own warning not to build anything load-bearing on it yet. It passed 160,000 GitHub stars within days, making the harness — not the model — DeepSeek's most-starred release.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: partly verified · confidence 70/100
- thenewstack.io News reporting archived
- cryptobriefing.com News reporting archived
- github.com Code repository archived
- deepseek.com Official documentation archived
-
Sakana AI launched Fugu Max and Fugu Ultra v2, an orchestration layer served through an OpenAI-compatible API that routes each task to the leanest model able to handle it, drawing on a pool of open-weights and specialised models including NVIDIA's Nemotron family. Sakana states the two are not separate products but two axes of one strategy: Max pushes cost efficiency, Ultra v2 pushes maximum capability.
Recorded as a milestone rather than a model because it orchestrates models rather than being one. Sakana calls Max and Ultra v2 models available through its API, but the post describes them routing work across a pool of other models — so a model record would carry a null parameter count, a null context window and a null licence, which is the shape that put ChatGPT and Bard here rather than in the release dataset.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 80/100
-
TypeSafe AI released Jev, which it calls its flagship model and the first "System One" model. Instead of generating text one token at a time, Jev takes a state and a set of typed questions and returns every answer in parallel as calibrated probabilities with confidence scores, for software to act on directly. It is served through an API and priced per input token; no weights were released.
A commercial model that does not generate text. Language models answer by producing tokens, and every structured output built on them is text parsed back into data; Jev answers typed questions directly with probabilities, in one parallel pass. Recorded as a milestone rather than a model release because this dataset's model types describe what a model produces, and none of them describes a decision. Whether the approach lasts is not yet known, so it is unranked.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 80/100
- typesafe.ai Official announcement archived
- docs.typesafe.ai Official documentation archived
-
OpenAI introduced dots, always-on agents powered by GPT-6 Astra, at DevDay 2026. Each has its own cloud computer, learns from feedback, connects to over 4,000 apps through plugins and is reached through ChatGPT, Slack or Teams. Rolled out first to Pro, Business Premium and Enterprise plans in eligible markets.
OpenAI's version of the persistent agent, seven weeks after xAI's Grok Bot. Like Grok Bot it runs on a model already recorded here (GPT-6 Astra), so it is an agent milestone rather than a model release.
No model line is recorded on this milestone, so nothing here links it to a model record.
Record status: verified · confidence 80/100
- openai.com Official announcement archived