AI agents are doubling in power every 3 months

Hi Futurist,

New research shows that AI agents are doubling in capability every three months. That’s not a headline, it’s a warning. New tools, smarter agents, sharper models are dropping weekly. If you're not paying attention, you're already behind. Therefore, I am sending you insights, inspiration, and innovation straight to your inbox. Let’s dive into the depths of the digital future together and discover the waves of change shaping our industry.

💡 In this post, we're dipping in:

• 📣 Byte-Sized Breakthroughs: AI agents are leveling up. And fast—they’re doubling in power every three months. With OpenAI’s new Agents SDK and Responses API, building your own AI workforce is easier than ever. Google dropped a new model, and Gemini got some serious upgrades.

• 🤖 Digital Toolbox: Icon turns your video library into ready-to-run ad campaigns using AI scene tagging and tools like AdGPT and AdCut. Relay blends smart automation with a human touch—connecting 100+ apps so your workflows flow without going rogue. And Ocoya speeds up social media creation with AI writing, scheduling, and cross-platform posting.

• 🧐 In Case You Missed It: OpenAI went full throttle with a lot of new update. Meanwhile, Google’s Gemini gets mind maps, YouTube link support, and real-time meeting notes. Adobe, Claude, Stability, and NVIDIA are all making major moves—from immersive 3D and image editing to real-time AI robotics.

Do you have tips, feedback or ideas? Or just want to give your opinion? Feel free to share it at the bottom of this dip. That's what I'm looking for.

No time to read? Listen to this episode of Digital Dips on Spotify and stay updated while you’re on the move. The link to the podcast is only available to subscribers. If you haven’t subscribe already, I recommend to do so.

Quick highlights of the latest technological developments.

Headstory: AI agents are doubling in power every 3 months

2025 is the year of the AI agents. That’s no longer a bold claim. It’s just reality. Agents are no longer just a tech demo on X or a promise in a product roadmap. They’re here. They write code. They draft emails. They answer customer support tickets. They fill in forms. And now they’re starting to string all those actions together, on their own.

Agents are learning to take more of the work off our plates. Not just the simple stuff. Not just one-shot tasks. But tasks with multiple steps. Tasks that require memory, planning, reasoning. In other words: real work. And they’re getting better at it fast.

New research from METR makes this clear. They’ve been tracking the kind of work agents can do reliably. Not “Can they solve this quiz question?” But: “Can they do the whole job?” Their method is simple. Measure how long it takes a skilled human to finish a task. Then see how often an AI can succeed at tasks of that length. That’s it. One number. They called it: the time horizon.

Back in the GPT-2 era, models couldn’t even complete a one-minute task. Today? Claude Sonnet 3.7 complete software tasks that take expert programmers 59 minutes on average, with a 50% success rate. In six years, AI time horizons have doubled every seven months. And now that’s speeding up: 2024 to 2025 models are doubling every three. Follow that curve, and by 2029, agents may be able to independently complete projects that take a skilled human a month.

But this makes it even astonishing: More than 80% of successful AI task runs cost less than 10% of what it would cost a human engineer to do the same work. Read that again. This isn’t just about efficiency anymore. This is a redefinition of economic value. A reshuffling of what labour is worth, and who—or what—should do it.

Let’s be clear: we are heading toward a world where work looks very different. This is not about replacing everyone with AI. This is about the shift from execution to delegation. From doing to deciding. From being the one doing the task… to becoming the one who assigns, reviews and corrects it. Everyone becomes a model manager. Everyone becomes a CEO of their own AI workforce.

Harvard recently conducted a study with 776 professionals at consumer goods giant Procter & Gamble (P&G) to assess AI’s impact on individuals and teams in developing products and retail strategies. The result? Individuals using AI already outperformed entire teams without it—and came close to matching the output of full teams with AI. Extend the current trend a few years out, and the picture sharpens: by 2029, one person could run an entire organization by themselves. The unit of productivity becomes the individual + their agents.

And yet, many still think this is a problem for “later.” But here’s the thing: the biggest mistake in any disruptive cycle is assuming today’s limits will hold tomorrow. They won’t.

So don’t wait. Start now. Build your AI layer. Learn how to work with AI agents. Use tools like Zapier, Relevance AI, Lindy, or Relay. Build your own AI workforce. Test. Fail. Improve. Repeat. Your job isn’t just to be better at your job. It’s to redesign how the job gets done. For leaders, it means rethinking team structures, workflows, hiring, value chains. For professionals, it means learning how to direct, review and collaborate with non-human teammates. Because in the future, success won't depend on how hard you work. It will depend on how well you manage your models.

OpenAI releases new tools to help organizations build their own AI agents

TL;DR

OpenAI introduces the Responses API, a new system to help developers build more useful and reliable AI agents. It bundles tools like web search, file search and computer use into one framework. The new Agents SDK lets you orchestrate and monitor multi-agent workflows. With this update, OpenAI simplifies the creation of agents that can handle complex, multi-step tasks.

Read it yourself?

Sentiment

People are excited about this release from OpenAI. Reactions like “There we go, these are some practical & powerful updates!” and “OpenAI is basically turning AI agents into real digital workers” are all over the timeline. The enthusiasm comes from the fact that OpenAI now offers tools that allow organizations to build agents fully embedded within their own systems. It’s a big step forward for enterprises.

My thoughts

Multi-agent systems are going to change how AI handles information. The shift isn’t just in smarter agents, but in how they work together. Pulling in data, checking sources, automating workflows. These updates mean developers and low-coders alike can now create agents that adapt to the organization’s context and data. And because the tools are now accessible at a relatively low cost, the door is open for everyone to experiment. By the end of the year, the real winners will be the organizations that manage to integrate agents into real processes, especially with humans-in-the-loop. Because that’s the future: not linear agent tasks, but dynamic systems. Of course, OpenAI had to make this move after the release of Manus (featured in the previous Dip). But here’s the key difference: Manus doesn’t use company-specific data and context. These new OpenAI tools do. Which means agents can now go deeper into actual workflows and become part of the real operational layer inside organizations.

Google releases Gemma 3: small size, big impact

TL;DR

Google DeepMind has launched Gemma 3, a new family of small but powerful open-source AI models. Built from the same tech behind Gemini 2.0, these models are fast, lightweight and designed to run on a single GPU or TPU—from phones to workstations. With support for 140 languages, vision-language tasks, and a 128k-token context window, Gemma 3 is proving you don’t need size to deliver serious results.

Read it yourself?

Sentiment

People are impressed. Not just because it’s open source, but because Gemma 3 delivers top-tier performance at a fraction of the size. The reaction? “Wow, Google keeps pushing on open weights AI in a big way.” And: “Pretty insane that new models today can perform better than leading models 15X the size just a year ago.” The big standout? Its massive context window and the fact it performs remarkably well as a non-reasoning model.

My thoughts

Gemma 3 tells us that high-performance AI is no longer reserved for those with massive compute clusters and enterprise budgets. What Google has done here is bring the power of modern language and vision models to a much smaller footprint, small enough to run on a single GPU or even locally on a laptop. This matters. A lot. Because the future of AI won’t only live in the cloud. Fast. Private. Always available. It will live in our phones, in our browsers, on the edge of networks. It will run without an internet connection. And it will still be smart enough to understand, reason, and act. Gemma 3 is proof that small models are catching up. And it beats models from just a year ago that were 15 times its size, but does so with fewer resources, less energy, and more accessibility. And that makes it deployable anywhere, from developing regions with limited infrastructure to embedded systems in industrial use cases. We’re entering an age where “small” is the new powerful. And with models like Gemma 3, that power is finally in everyone’s hands.

More byte-sized breakthroughs:

• Baidu launches two AI models at half the cost of its rivals
  China’s Baidu just released ERNIE 4.5 and ERNIE X1—two new AI models now free for individual users. ERNIE X1 performs on par with DeepSeek R1, but costs half as much. ERNIE 4.5 improves reasoning, coding, and hallucination control, and is already live via API. While OpenAI is eyeing $20,000+ price tags for agents, Baidu is slashing costs and pushing the market toward mass access. The AI arms race just got cheaper.

• OpenAI launches smarter, more customizable audio models
  OpenAI just rolled out new speech-to-text and text-to-speech models with major upgrades in accuracy and control. The models handle accents, noise, and varying speech speeds better than ever. Developers can now build voice agents that don’t just sound human, but feel human. You can even instruct the voice to act like a support rep or a bedtime storyteller. It's not just speech recognition anymore. It’s conversation on your terms.

• Canvas makes Gemini the most hands-on AI yet
  Gemini’s new Canvas feature just changed the AI game overnight. You can now create interactive prototypes, write docs, and design landing pages—all in real time, with AI edits as you go. Need a break from reading? Audio Overview turns your files into podcast-style conversations. Share your work via subdomains and let others jump in. Gemini is now outpacing ChatGPT and Claude in speed, style and usefulness.

A must-see webinar, podcast, or article that’s too good to miss.

Icon - Create winning ads in minutes

Icon is an AI-powered ad maker that helps you create effective ads quickly. It analyses your video library, tags scenes, and uses these clips as building blocks for new ads. With features like AdGPT and AdCut, you can streamline your ad creation process and reduce production time and costs.

Relay - Automate workflows with AI and human touch

Relay.app is a modern automation tool that connects over 100 apps, including Gmail, Slack, and Notion. It combines AI capabilities with human-in-the-loop features, allowing you to automate tasks while retaining necessary oversight. With deep integrations and a user-friendly interface, it simplifies your workflow.

Ocoya - Create social media content faster

Ocoya is an AI-powered platform that helps you generate and post content quickly, both manually and automatically. It combines AI writing, automation, scheduling, and analytics to streamline your social media management. With integrations to over 30 platforms, Ocoya simplifies your content creation process.

A roundup of updates that are too cheesy to ignore.

• OpenAI unveils a 'new model' adept in creative writing, as per CEO Sam Altman’s announcement, marking a shift towards storytelling.

• OpenAI introduces Python-powered data analysis with o1 and o3-mini in ChatGPT, enabling tasks like regression analysis and metric visualization.

• OpenAI updated their Playground, designed for testing, comparing, and iterating on prompts with new tools like web and file search.

• OpenAI's Work with Apps is now available on macOS for all users, enabling ChatGPT to read and edit content in your coding apps.

• OpenAI’s Operator Update is now available to all Pro users in the EU, Switzerland, Norway, Liechtenstein, and Iceland.

• OpenAI will soon test ChatGPT Connectors, allowing business users to integrate Slack and Google Drive into the AI chatbot environment.

• OpenAI’s o1-Pro is now available through their API, offering enhanced compute for superior responses, with vision and structured outputs, at a premium cost.

• Harvey's new Workflows launch today, transforming legal work with adaptive, reasoning agents that rival human lawyers in blind tests.

• MoveAI introduces advanced 3D motion capture, full-body dynamics, and motion prediction.

• Granola's new chat feature enables seamless communication with any contact, eliminating the need to sift through past notes.

• The EU releases a new draft of the AI Act Code of Practice, aiming for clarity and compliance as the May deadline approaches.

• Stability AI unveils Stable Virtual Camera, turning 2D images into immersive 3D videos with realistic depth and complex reconstructions.

• HeyGen's Brand Kit streamlines video creation by instantly pulling your colors, fonts, and logos from your domain.

• Pika's new Manipulation Feature lets you alter any character or object in your video, leaving everything else untouched.

• Krea uses Google’s Veo 2, the most powerful video generation model, to seamlessly work with start images.

• Krea's introduces Video Training by allowing you to train Wan 2.1 with custom videos to learn unique styles, motions, or objects.

• LumaLabs unveils a new pre-training breakthrough offering superior sample quality and 10x efficiency for diffusion models.

• Nex releases Ikon-2, offering unmatched quality and efficiency in image generation with up to 8MP photos for stunning product consistency.

• Freepik launches Retouch, powered by Gemini, allowing you to enhance your images with simple prompts for stunning results.

• Freepik introduces Objects, enabling the addition of the same item across multiple images seamlessly.

• Freepik now includes Elements, letting you integrate reference images into AI videos to ensure consistency across elements and characters.

• Bolt introduces seamless integration with Figma, transforming your designs into pixel-perfect full-stack apps with a simple URL prompt.

• Lovable introduces Custom Domains, allowing you to seamlessly connect a personalized domain to your app within the platform.

• Convergence introduces DeepWork, their most sophisticated AI agent yet, autonomously coordinating multi-agent workflows.

• Cohere introduces Command A, an efficient new model excelling in agentic tasks with minimal compute supporting 23 languages.

• Google enhances Deep Research in GeminiApp with the Gemini 2.0 Flash Thinking Experimental, now open for all users to explore insights at no cost.

• Google's Gemini introduces striking improvements, including extend context windows, personalized interactions, and connected apps, all free.

• Google AI Studio introduces support for YouTube video links via the Gemini API, enabling native video understanding with just a link.

• Google Meet's Gemini now takes meeting notes for you, allowing you to focus fully on the meeting.

• Google acquires cloud security startup Wiz for $32B to strengthen its cloud security offerings.

• Google introduces interactive Mindmaps in NotebookLM, transforming notebooks into personalized learning experiences.

• Xbox's Copilot for Gaming helps players improve skills with AI-driven assistance coach, guiding you through games.

• ServiceNow’s Yokohama Platform redefines enterprise AI by unifying automation and intelligence with new AI agents and workflow solutions.

• Apple unveils a new AirPods feature in iOS 19 for real-time conversation translation, enhancing user experience akin to Google Pixel Buds.

• The US AI Alliance shifts focus as scientists are guided to eliminate skills like AI safety, responsible AI, and AI fairness from advanced models.

• Grok's new update lets anyone edit any image just by talking, making image editing incredibly simple.

• xAI acquires Hotshot, a generative AI video startup bolstering their capabilities in video creation technologies.

• Researchers unveil QNodeOS, the first OS for quantum networks, which aims to make quantum internet a reality.

• LG AI Research unveils EXAONE Deep, setting new benchmarks in reasoning AI with superior performance across math, science, and coding.

• NVIDIA introduces Blue, a Star Wars-inspired robot featuring a new physics engine with real-time intelligence and movement.

• NVIDIA introduces the DGX Spark and DGX Station, empowering users to develop and refine AI models at the edge.

• NVIDIA has reportedly purchased Gretel, a startup specializing in synthetic AI training data, marking a strategic expansion in AI capabilities.

• Adobe launches their Agent Platform that enables businesses to activate AI agents for enhanced customer experiences and marketing workflows.

• Ripple celebrates as the SEC decides to drop its appeal—a triumphant win for the crypto community worldwide.

• Claude now includes web search to enhance its responses with up-to-date, verified, real-time information. Available in the US only, for now.

• Mistral AI introduces Small 3.1, a versatile and speedy multimodal model with advanced customization options for on-device use.

• Cloudflare introduces AI Labyrinth, a new tool that uses AI-generated pages to disorient unauthorized crawlers and identify bot activity.

This was it. Our thirty-second digital dip together. It might seem like a lot, but remember; this wasn't even everything that happened in the past few weeks. This was just a fraction.

If you're serious about staying ahead, I can help. Whether you're a professional, organization or leading a team: let’s build your AI workforce together. I guide organizations through this shift, from setup to strategy. Just hit reply, and I’ll drop by.

Looking forward to what tomorrow brings! ▽

-Wesley