General AI agents
2 labs · 15 models and products Theme: Companies · Generated from this page's data. Click a name to open its panel on the Labs page. Records resting on one source are tagged "one source".An AI agent is a system that does not just answer questions but carries out tasks: clicking through websites, playing games, or writing and fixing code. Most agents on this tile come from Google DeepMind: Gato, one network that plays games, captions images and controls a robot arm; SIMA, which follows instructions in 3D video games; and CodeMender, which finds and fixes security flaws (per the site). Convergence, a London start-up, built Proxy, a personal web agent (per the site); Salesforce bought it in 2025 and said it would build a centre of AI research in London.
Context
- DeepMind's Gato showed one AI model handling many kinds of task, input and body (simulated or robotic). 12 May 2022 arxiv.org/abs/2205.06175
The agent, which we refer to as Gato, works as a multi-modal, multi-task, multi-embodiment generalist policy.
- DeepMind's Agent57 was the first deep reinforcement-learning agent to beat the standard human benchmark on all 57 Atari games. 30 March 2020 arxiv.org/abs/2003.13350
We propose Agent57, the first deep RL agent that outperforms the standard human benchmark on all 57 Atari games.
- Google DeepMind's SIMA team set out to build an agent that follows instructions in any simulated 3D world. 13 March 2024 arxiv.org/abs/2404.10179
develop an instructable agent that can accomplish anything a human can do in any simulated 3D environment
- Salesforce agreed to buy Convergence on 15 May 2025 (closed 11 June 2025) and said the deal would found an AI research centre in London. 15 May 2025 salesforce.com/news/stories/salesforce-signs-def
lay the foundation for a growing center of AI R&D here
- The AI Security Institute reports that some abilities AI agents would need to evade human control are improving. 18 December 2025 aisi.gov.uk/blog/5-key-findings-from-our-first-f
Some of the capabilities that would be required for AI models to evade human control are improving
Key papers
- A Generalist Agent (Gato) arXiv preprint · 12 May 2022
- Scaling Instructable Agents Across Many Simulated Worlds (SIMA) arXiv preprint · 13 March 2024
- Agent57: Outperforming the Atari Human Benchmark arXiv preprint · 30 March 2020
Labs Round 8 (31 Oct 2026): introduction and facts as the researchers wrote them; "(per the site)" marks a sentence resting on this site's own description of that lab.
| Record | What it is | Place | Links | Confidence |
|---|---|---|---|---|
| Agent57Chips, models and products | A DeepMind game-playing agent reported to outperform the human benchmark on Atari games. released (paper)Key facts Theme: General AI agents. Blog March 2020 (index title). Details page not opened.Made by Google DeepMind · 2020-03 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/page/14/ | one source |
| AI co-clinicianChips, models and products | Google DeepMind research towards an AI that works alongside clinicians in patient care. researchKey facts Theme: General AI agents. Post dated 30 Apr 2026.Made by Google DeepMind · 2026-04-30 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/rss.xml | one source |
| ALOHA Unleashed and DemoStartChips, models and products | Two robot-learning methods: one teaches two-armed robots complex tasks, the other trains dexterous hands in simulation. research (paper only)Key facts Announced 12 Sep 2024; DemoStart needs 100x fewer simulated demonstrations; over 98% success in simulation on several tasks; real-world 97% on cube reorientation and lifting, 64% on plug-socket insertion; access: paper onlyMade by Google DeepMind · 2024-09-12 | London | deepmind.google/blog/advances-in-robot-dexterity | one source |
| Co-ScientistChips, models and products | A Gemini-based multi-agent AI partner built to help researchers speed up scientific work. releasedKey facts Theme: General AI agents. Announced 12 May 2026; same week as "Gemini for Science" tools (17 May 2026) and five science case posts (16 May 2026).Made by Google DeepMind · 2026-05-12 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/rss.xml · deepmind.google/blog/rss.xml | lab's own pages |
| CodeMenderChips, models and products | An AI agent that finds and fixes security flaws in software and rewrites code to make it safer. research (not public; human-reviewed patches)Key facts Theme: General AI agents. Announced 6 Oct 2025. Figures (6 Oct 2025): "72 security fixes to open source projects", some in codebases as large as 4.5 million lines. Not publicly available; patches reviewed by people before submission.Made by Google DeepMind · 2025-10-06 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/introducing-codemender-an-a · deepmind.google/blog/page/5/ | one source |
| ConvergenceCompanies | London start-up that built Proxy, a personal AI agent that learns and carries out multi-step tasks on the web; now part of Salesforce. acquiredProfile
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: verified. Sources | London | salesforce.com/news/stories/salesforce-signs-def · tech.eu/2025/05/15/salesforce-acquires-uk-ai-sta · tech.eu/2024/09/27/convergence-secures-12m-pre-s · salesforce.com/news/stories/salesforce-signs-def | verified |
| CursiveCompanies | Lab founded by senior DeepMind alumni building AI agents that keep learning from real-world use; described as a foundation model start-up.Profile
Labs Round 7 and Labs Round 8 deep profile (30–31 Oct 2026), as the researchers wrote it; confidence: one source (nothing verified) ---. Sources Key facts and news (3)
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | gov.uk/government/news/ai-firms-pioneering-drug- · tech.eu/2026/05/01/over/ · sovereignai.gov.uk/post/our-first-investments · telecomtv.com/content/digital-platforms-services | verified | |
| GatoChips, models and products | One neural network with one set of weights that plays games, captions images, chats and controls a robot arm. released (paper)Key facts Theme: General AI agents. Announced 12 May 2022. Figure stated: context window of 1,024 tokens. Access: research publication; no product.Made by Google DeepMind · 2022-05-12Read the paper | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/a-generalist-agent/ · deepmind.google/blog/page/12 | one source |
| Gemini 2.5 Computer UseChips, models and products | A Gemini model that operates computer interfaces such as web pages by clicking and typing. released (model card)Key facts Theme: General AI agents. Announced October 2025 (index); model card updated 7 Oct 2025. Later: computer use in Gemini 3.5 Flash (24 Jun 2026).Made by Google DeepMind · 2025-10 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/page/5/ · deepmind.google/models/model-cards/ | lab's own pages |
| Google Antigravity (and Antigravity 2.0)Chips, models and products | A Google product, launched November 2025, through which Gemini models and DeepMind tools (such as an AlphaGenome skill) are offered; version 2.0 followed in May 2026. released (product)Key facts Theme: General AI agents. "Introducing Google Antigravity" November 2025 (index); "Introducing Google Antigravity 2.0" 17 May 2026; used for antimicrobial peptide search (15 Oct 2026) and as a channel for Gemini models (model cards).Made by Google DeepMind · 2025-11 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/page/5/ · deepmind.google/blog/rss.xml | lab's own pages |
| Open X-Embodiment dataset and RT-X models (RT-1-X, RT-2-X)Chips, models and products | Pooled robot-training dataset from many labs and the robot models trained on it, which work across different robot types. released (open dataset; models)Key facts Announced 3 Oct 2023; data from 22 robot types; more than 500 skills, 150,000 tasks, more than 1 million episodes; partners from 33 academic labs; RT-1-X outperformed the original model by 50% on average; access: open datasetMade by Google DeepMind · 2023-10-03 | London | deepmind.google/blog/scaling-up-learning-across- · deepmind.google/blog/2023-a-year-of-groundbreaki | lab's own pages |
| Project AstraChips, models and products | Google DeepMind's prototype universal assistant that sees and hears through a phone camera or prototype glasses and talks back in real time. research prototype (trusted testers)Key facts Used by a limited number of trusted testers (waitlist); version for blind and low-vision users in a trusted-tester programme; runs on Android phones and prototype glasses; features such as screen sharing and video understanding moved into Gemini LiveMade by Google DeepMind · — (date not stated on the project page) | London | deepmind.google/models/project-astra/ | one source |
| PROWL-1 and PROWL-2Chips, models and products | Odyssey research models that learn to act inside simulated worlds; the second learns simulation and decision-making together. announced (research)Key facts PROWL-1 announced 12 May 2026 ("Learning Through Discovery"); PROWL-2 announced 1 Oct 2026; posts not opened (links blocked)Made by Odyssey · 2026-05-12 | odyssey.systems/writing | one source | |
| RoboCatChips, models and products | Self-improving robot agent that learns new tasks on different real robot arms from a few hundred demonstrations. research (paper only)Key facts Announced 20 Jun 2023; learns a new task from 100 to 1,000 demonstrations; data from four robot types; success on unseen tasks after 500 demonstrations rose from 36% (initial) to 74% (final); arXiv 2306.11706; access: paper onlyMade by Google DeepMind · 2023-06-20 | London | deepmind.google/blog/robocat-a-self-improving-ro | one source |
| SIMA (Scalable Instructable Multiworld Agent)Chips, models and products | An AI agent that follows spoken-style instructions to carry out tasks in 3D video games using keyboard and mouse. research only (technical report)Key facts Theme: General AI agents. Announced 13 Mar 2024 as early-stage research. Figures (13 Mar 2024): nine video games; eight game studio partners; 600 basic skills. Successor: SIMA 2 (13 Nov 2025).Made by Google DeepMind · 2024-03-13 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/sima-generalist-ai-agent-fo · deepmind.google/blog/page/9/ | one source |
| SIMA 2Chips, models and products | A Gemini-powered game agent that follows instructions, reasons about goals, talks with users and improves itself in 3D worlds. limited research previewKey facts Theme: General AI agents. Announced 13 Nov 2025; early access for a small group of academics and game developers. No numeric success rates in text. Predecessor SIMA (Mar 2024) learned over 600 skills.Made by Google DeepMind · 2025-11-13 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/sima-2-an-agent-that-plays- · deepmind.google/models/ | lab's own pages |
| XLand open-ended learning agentsChips, models and products | DeepMind agents trained across hundreds of thousands of generated 3D games that learned general game-playing behaviour. research (paper only)Key facts Announced 27 Jul 2021; about 700,000 unique games across 4,000 unique worlds; final agents trained on 3.4 million unique tasks; access: paper onlyMade by Google DeepMind · 2021-07-27 | London | deepmind.google/blog/generally-capable-agents-em | one source |