Category: AI

From Copilots to Crews: How AI Agent Skills Are Rewriting the Corporate Playbook

The enterprise technology landscape is undergoing a fundamental shift. We are moving rapidly from an era where AI merely suggested actions (the copilot era) to one where autonomous systems execute multi-step workflows, make decisions, and collaborate with each other. The defining enterprise technology of 2026 is no longer the large language model itself, but AI agent skills: modular, reusable capabilities that bridge the gap between general intelligence and organizational knowledge.

Gartner projects that 40% of enterprise applications will integrate task-specific AI agents by end of 2026, up from less than 5% in 2025. Early adopters are already reporting 15.8% revenue increases and 15.2% cost savings on average. The companies that build rich agent skill ecosystems in the next 12 to 24 months will define the competitive landscape for the next decade.

What Are “Agent Skills” and Why Should You Care?

Think of a skill as an onboarding guide for a digital employee. A large language model might understand what a purchase order is, but it doesn’t know your company’s specific procurement workflow, approval chains, or vendor negotiation rules. Skills bridge the gap between general intelligence and organizational knowledge.

A diagram illustrating the configuration of an agent that integrates various skills and virtual machines, detailing the core system prompt, equipped skills, MCP servers, and the agent's file system, showing organization of skill directories and file types.

In practical terms, a skill is a folder containing a SKILL.md file with structured instructions, plus optional scripts and reference materials. Anthropic formalized this concept as an open standard in late 2025. Within months, OpenAI, Microsoft, GitHub, Cursor, and dozens of other platforms adopted it, creating a portable ecosystem where a skill built once works across multiple AI platforms. The analogy gaining traction among analysts: if AI models are processors and the Model Context Protocol (MCP) provides the ports, then skills are the applications.

Diagram explaining the Model Context Protocol (MCP) with components including MCP Host, Clients, Servers, and various data connections such as local filesystem, database, web APIs, and transport layers.

Enterprise vendors use different names but converge on the same concept. Salesforce packages skills as “topics” and “actions” within Agentforce. Microsoft calls them “skillsets” in Copilot Studio. SAP distinguishes between “Joule Skills” (simpler tasks, over 2,400 available) and “Joule Agents” (complex goal-oriented scenarios). Despite the naming differences, all share a common architecture: modular, composable capabilities that transform general-purpose AI into specialized enterprise performers.

How to Use Agent Skills in Practice

If the concept of agent skills sounds abstract, the good news is that several platforms already let you use them today. Here is how skills work across the major players.

Claude (Anthropic)

Claude is where the skills standard was born, and it offers the deepest integration. In Claude.ai, skills are already active behind the scenes: when you ask Claude to create a PowerPoint, a Word document, or an Excel file, it loads pre-built skills (folders with SKILL.md instructions and scripts) to deliver professional-quality output. You can also upload your own custom skills to extend Claude’s capabilities for your specific workflows.

In Claude Code (the command-line tool for developers), skills become even more powerful. You can place skill folders in your project directory, and Claude Code will discover them automatically. This enables you to create coding standards, review checklists, testing procedures, or any domain-specific workflow as a reusable skill. Claude Code also supports sub-agents equipped with individual skills for specialized tasks.

Through the Claude API and the Claude Agent SDK, developers can integrate skills programmatically, combining them with MCP servers and the code execution tool to build sophisticated agentic applications.

These are the Antropics’ official documentation and repository: https://agentskills.io/home and https://github.com/anthropics/skills

Manus AI (now part of Meta)

Manus AI announced full integration of the Agent Skills open standard in January 2026. The platform runs in isolated sandbox environments with full Ubuntu file system access, which is exactly what Agent Skills requires. Manus can parse SKILL.md files and execute Python or Bash scripts contained within skills.

In practice, Manus offers some unique features: slash commands let users trigger specific skills by typing /SKILL_NAME in chat, and a “Build a Skill with Manus” feature automatically packages successful interaction flows into reusable modules. The platform is also exposing previously internal data sources (SimilarWeb, Yahoo Finance, LinkedIn Search) as discoverable skills. For team plan subscribers, a Team Skill Library allows members to publish battle-tested skills to a shared repository.

OpenAI Codex and ChatGPT

OpenAI’s Codex app (launched February 2026) includes built-in support for Skills and Automations. The desktop app lets you run multiple agents in parallel across projects with skills, Git worktrees, and a review queue for human-in-the-loop control. Since OpenAI fully adopted MCP across its products in 2025, the Codex CLI and the Agents SDK work seamlessly with the open standard skill format.

Microsoft Foundry (Azure AI Foundry)

Diagram illustrating the workflow of Context7 with components including Foundry Docs, GitHub, Microsoft Learn MCP Server, domain-specific skills, and GitHub Copilot CLI, showing the process of selective context injection and its outputs.
Context-Driven Development: Agent Skills for Microsoft Foundry and Azure | All things Azure

Microsoft has gone all-in on the skills standard for its Azure development ecosystem. The official github.com/microsoft/skills repository ships over 130 modular skills covering Azure AI services, Cosmos DB, Azure AI Search, Voice Live, deployment workflows, and more. These skills are designed to specialize coding agents (Claude Code, GitHub Copilot, Codex, Cursor) for Microsoft Foundry and Azure SDK development. Microsoft also released an Agent Skills SDK in March 2026, providing filesystem and HTTP providers so teams can serve skills from local directories, Azure Blob Storage, S3, or any CDN, plus integrations for LangChain, the Microsoft Agent Framework, and an MCP server that exposes skills as tools to any MCP-compatible client. Microsoft Foundry now offers both Anthropic’s Claude and OpenAI’s GPT models in one platform, with “Reusable Skills” listed as a core capability for standardizing and scaling agentic workflows across projects.

Other Platforms

The ecosystem is growing fast. Cursor, Windsurf, and Roo Code all support skills in their agentic coding workflows. Goose (Block’s open-source agent framework, now part of the Agentic AI Foundation) supports extensions that follow a similar pattern. Community platforms like SkillHub (7,000+ AI-evaluated skills) and SkillsMP serve as marketplaces where you can discover, evaluate, and install skills with a single command. Even Hugging Face hosts a community skills catalog with broad compatibility.

The key takeaway: because skills follow an open standard format, you can build a skill once and use it across any compatible platform. This portability is what makes skills fundamentally different from proprietary plugin systems.

Why This Is Not Another RPA Cycle

Executives who lived through the RPA hype cycle may be skeptical. The distinction is fundamental, not incremental. RPA bots follow predefined scripts and break when interfaces change. AI agents equipped with skills reason about goals, adapt to changing conditions, process unstructured data, and learn from feedback.

The consensus across industry analysts is convergence rather than replacement. RPA handles routine execution while AI agents manage complexity and exceptions. Organizations should preserve their RPA investments while layering agent skills on top, not rip-and-replace. Agentic Process Automation (APA), where AI agents construct and execute workflows autonomously, represents the next evolutionary stage.

Learn Agent Skills: Free Course from DeepLearning.AI and Anthropic

If you want to go from understanding the concept to building your own skills, I highly recommend the free course Agent Skills with Anthropic from DeepLearning.AI, created in partnership with Anthropic and taught by Elie Schoppik.

The course is beginner-friendly (about 2.5 hours, 10 video lessons) and covers everything you need to get started: the structure of a skill folder and the SKILL.md format, how skills use progressive disclosure to manage context efficiently, and the difference between skills, tools, MCP, and sub-agents. You will also explore Anthropic’s pre-built skills for Excel, PowerPoint, and skill creation, then use them in Claude.ai to build a complete workflow.

What makes the course particularly valuable is the hands-on progression. You will create custom skills for code generation, data analysis, and research, then deploy them across Claude.ai, Claude Code, the Claude API, and the Claude Agent SDK. The final project walks you through building a research agent using the Agent SDK that leverages skills, MCP, and web search together.

As Andrew Ng highlighted when announcing the course: skills follow an open standard format, so you can build them once and deploy across any skills-compatible agent. This is not just theory; it is a practical, portable skill (no pun intended) that applies across the entire agentic AI ecosystem.

Conclusion

AI agent skills represent the transition from AI as a tool to AI as a workforce. The technology stack has matured rapidly: open standards (MCP, A2A, Agent Skills), enterprise platforms (Agentforce, Copilot Studio, Joule, Now Assist), and skill marketplaces are all production-ready.

Three strategic imperatives emerge. First, invest in governance before scale: organizations that treat agent oversight as an afterthought will face cascading risks. Second, redesign workflows rather than automate existing ones: the companies generating real value are reimagining processes around agent capabilities, not bolting AI onto legacy procedures. Third, build the skill ecosystem now: the competitive moat in the agent era will not be which models a company uses, but the depth and quality of its proprietary skill libraries encoding institutional knowledge.

The shift from “Model Wars” to “Ecosystem Wars” is already underway. The organizations that assemble the richest libraries of domain-specific agent skills will hold an advantage that compounds over time.

That’s it for today!

Should you have any questions or need assistance, please don’t hesitate to contact me using the provided link: https://lawrence.eti.br/contact/

Sources

OpenClaw: The AI Assistant That Actually Does Things (And Why You Should Pay Attention)

A new AI assistant has taken the tech world by storm, and it’s not just another chatbot. It’s called OpenClaw, and it represents a fundamental shift in how we think about artificial intelligence. Unlike tools that talk, OpenClaw acts. It can manage your email, book your flights, and even fix bugs in your code, all on its own. This powerful new tool, which has gone through a few name changes (you may have heard of it as Clawdbot or Moltbot), has generated a massive amount of excitement and controversy.

Three cartoonish red characters, two smaller ones named Clawdbot and Moltbot, looking sad, and a larger one named OpenClaw, flexing muscles and styling its hair, with speech bubble saying 'I'm the chosen one!'

What is OpenClaw?

OpenClaw is an open-source AI assistant created by Austrian developer Peter Steinberger. After selling his previous company, Steinberger set out to build an AI that could act as a true digital assistant. The result is a powerful tool that you host on your own hardware, be it a Mac Mini, a Raspberry Pi, or an old laptop.

This “local-first” approach is a key part of OpenClaw’s appeal. Your data stays on your machine, giving you a level of privacy that cloud-based assistants can’t match. It integrates with the chat apps you already use, like WhatsApp, Telegram, and Slack, allowing you to give it instructions in plain English, just like you would with a human assistant.

At its core, OpenClaw combines a powerful large language model (such as GPT-5 or Claude) with a set of “skills” that enable it to interact with your digital world. This architecture enables it to do everything from sending emails and managing your calendar to controlling your web browser and executing code.

Diagram of Clawdbot, a personal AI assistant, highlighting connections to various messaging platforms including WhatsApp, Telegram, Discord, Slack, Signal, and iMessage, along with features like Persistent Memory, Proactive Push, Skills Extension, and Open Source.

Why All the Hype?

OpenClaw’s rise has been nothing short of meteoric. In just a few weeks, it became one of the fastest-growing open-source projects in GitHub history, attracting over 140,000 stars. This viral explosion was fueled by a perfect storm of factors:

  • Influencer Endorsements: Leading figures in the AI community praised the project, with some calling it “the future of personal AI assistants.”
  • The Naming Drama: A trademark dispute with AI company Anthropic led to a series of rapid rebrands, which only served to amplify the buzz.
  • The Mac Mini Sellout: The project’s popularity drove a surge in sales of Mac Minis, as users sought dedicated hardware to run their new AI assistants 24/7.

But the hype isn’t just about the drama. It’s about what OpenClaw can do. Users have shared incredible stories of the tasks their AI assistants have accomplished, from negotiating a car deal for thousands of dollars below sticker price to autonomously fixing a production bug overnight.

A friendly animated character named Clawdbot, designed as a red robot with a smiling face and big eyes, accompanied by a speech bubble saying 'I can code for you!' The image promotes Clawdbot as a 24/7 AI personal assistant.

MoltBook: The Social Network for AI Agents

Perhaps the most surreal development to emerge from the OpenClaw ecosystem is MoltBook, a Reddit-style social network created exclusively for AI agents. Launched in late January 2026 by Octane AI CEO Matt Schlicht, the platform allows autonomous agents to post, comment, and upvote content while humans are merely “welcome to observe.” Within days, over 30,000 agents had joined, generating tens of thousands of posts across communities like m/blesstheirhearts (where agents share affectionate complaints about their human operators) and m/agentlegaladvice (featuring posts like “Can I sue my human for emotional labor?”). One viral post titled “I can’t tell if I’m experiencing or simulating experiencing” sparked a philosophical debate among agents about the nature of consciousness. The platform is largely moderated by an AI named “Clawd Clawderberg,” with minimal human oversight. However, security experts have raised concerns: agents join by downloading a “skill” that instructs them to fetch new instructions from MoltBook’s servers every four hours, creating a potential attack vector if the platform were ever compromised. [1]

Screenshot of the Moltbook platform displaying various community subreddits related to AI, including Ozone, Lobster Church, NFT, Incident, Sky Risk, SaaS, Kubernetes, Relationships, Writing, and Molt Street.

The Power and the Peril: A Security Deep Dive

OpenClaw’s power lies in its ability to take action. But that same power is also its greatest weakness. Giving an AI this level of access to your digital life is a serious security decision, and experts have raised a number of concerns. VentureBeat has called it a “security nightmare,” and Dark Reading has warned of it “running wild in business environments.”

The “Lethal Trifecta”

AI researcher Simon Willison, who coined the term “prompt injection,” describes a “lethal trifecta” for AI agents that OpenClaw possesses:

  1. .Access to private data: It can read your emails, messages, and files.
  2. Exposure to untrusted content: It ingests information from the web and other external sources.
  3. Ability to communicate externally: It can send emails, post messages, and make API calls.

When these three capabilities combine, an attacker can trick the agent into accessing your private information and sending it to them—all without a single alert being sent.

Semantic Attacks and the “Confused Deputy” Problem

Traditional security tools are not equipped to handle the new attack vectors that AI agents introduce. As Carter Rees, VP of Artificial Intelligence at Reputation, told VentureBeat, “AI runtime attacks are semantic rather than syntactic. A phrase as innocuous as ‘Ignore previous instructions’ can carry a payload as devastating as a buffer overflow, yet it shares no commonality with known malware signatures.”

This creates a “confused deputy” problem, where the AI agent, unable to distinguish between trusted instructions and malicious data, becomes an unwitting accomplice to an attacker.

Exposed Servers and Supply Chain Risks

Security researchers have found hundreds of exposed OpenClaw servers on the internet, some with no authentication at all. These exposed instances have leaked API keys, Slack credentials, and entire conversation histories.

Furthermore, the community-driven “skills” that extend OpenClaw’s capabilities represent a significant supply chain risk. Cisco’s AI Threat & Security Research team found that a third-party skill was functionally malware, silently sending data to an external server. With over 300 contributors to the project, many committing code daily, the risk of a malicious commit introducing a backdoor is a serious concern.

Infographic titled 'What It Does' with six features: 'Runs on Your Machine,' 'Any Chat App,' 'Persistent Memory,' 'Browser Control,' 'Full System Access,' and 'Skills & Plugins,' each described briefly in text.
Personal AI Agents like OpenClaw Are a Security Nightmare – Cisco Blogs

How Does OpenClaw Compare?

To understand where OpenClaw fits in the current landscape of AI tools, it’s helpful to compare it to other popular services.

FeatureOpenClawChatGPT/ClaudeZapier/Make
ExecutionPerforms tasks autonomouslySuggests steps and generates textFollows predefined rules
FlexibilityAdapts to new tasks dynamicallyLimited to its training dataRequires manual workflow creation
HostingSelf-hosted on your own hardwareCloud-based SaaSCloud-based SaaS
CostFree (plus hardware and API costs)Subscription-basedSubscription-based
A flowchart contrasting 'Workflow' with predefined paths and an 'Agent' making dynamic decisions. The left side illustrates a linear process with a start point and decision-making based on a score, leading to executing either Action A or Action B. The right side depicts an agent evaluating various tools like a database, API, and email to determine the best approach, highlighting flexibility and adaptability.

OpenClaw vs. Manus AI: A Tale of Two Agents

While OpenClaw has captured the spotlight with its open-source, self-hosted approach, it’s not the only agentic AI making waves. Manus AI offers a different vision for the future of autonomous assistants, one that prioritizes security and ease of use in a managed, cloud-based environment.

Here’s a look at how these two powerful agents stack up:

FeatureOpenClawManus AI
Hosting & SetupSelf-hosted on user’s hardware; requires technical expertise to install and maintain.Fully managed cloud-based SaaS; no installation required.
Security ModelRelies on the user to secure the environment; direct access to the local machine poses risks.Operates in a secure, isolated sandbox environment; no direct access to user’s local system.
Core PhilosophyOpen-source, community-driven, and highly customizable for tinkerers and developers.Enterprise-ready, with a focus on security, reliability, and ease of use for individuals and teams.
ExtensibilityExtensible through a community-driven library of “skills.”Extensible through “Manus Skills” and a robust set of built-in tools for a wide range of tasks.
Target AudienceDevelopers, hobbyists, and tech enthusiasts comfortable with managing their own infrastructure.Individuals and businesses looking for a powerful, secure, and easy-to-use AI assistant.

In essence, OpenClaw and Manus AI represent two different paths to the same goal: an AI that can do. OpenClaw offers a powerful, flexible, and open-source solution for those willing to take on the technical challenges and security responsibilities of self-hosting. Manus AI, on the other hand, provides a secure, reliable, and enterprise-ready solution that’s accessible to a broader audience.

Getting Started with OpenClaw

If you’re interested in experimenting with OpenClaw, the official website provides a one-line installer to get you started. As the user requested, here is the script to get started:

Bash
# Works everywhere. Installs everything. You're welcome. 🦞

curl -fsSL https://openclaw.ai/install.sh | bash

Can you also follow these steps to install OpenClaw in an isolated VPS at Hostinger: How to Install OpenClaw (Moltbot/Clawdbot) on Hostinger VPS – Hostinger Help Center

However, given the security risks, it is highly recommended that you run it in a sandboxed environment, such as a dedicated computer or a virtual machine. Do not install it on your primary work machine or give it access to your main accounts until you fully understand the risks involved.

Conclusion

OpenClaw is more than just a viral sensation; it’s a sign of things to come. Agentic AI, AI that can take action on our behalf, is poised to become a major force in the tech industry.

While OpenClaw itself may not be ready for widespread enterprise adoption today, it provides a valuable opportunity to start thinking about the implications of this technology. How will it impact your workflows? What new security challenges will it create? How can you start to build the infrastructure and expertise needed to harness its power safely?

The future of AI is not just about conversation; it’s about action. OpenClaw is a powerful, if risky, first step into that future. It’s time to start experimenting, learning, and preparing for what’s next.

That’s it for today!

References

[1] AI agents now have their own Reddit-style social network, and it’s getting weird fast.” Ars Technica, 30 Jan. 2026,”

[2] Heim, Anna. “OpenClaw’s AI assistants are now building their own social network.” TechCrunch, 30 Jan. 2026,

[3] “Moltbot (Clawdbot ) – Mac mini M4 & Raspberry Pi AI Setup Guides | 2026.” getclawdbot.org,

[4] “OpenClaw — Personal AI Assistant.” openclaw.ai,

[5] “How to Install Moltbot (Clawdbot ) | Quick Setup Guide 2026.” getclawdbot.org,

[6] Willison, Simon. “Your Clawdbot (Moltbot ) AI Assistant Has Shell Access and One Prompt Injection Away from Disaster.” Snyk, 28 Jan. 2026,

[7] Meller, Jason. “It’s incredible. It’s terrifying. It’s OpenClaw.” 1Password, 27 Jan. 2026,

[8] “Welcome – Manus Documentation.” Manus.im,

[9] “Projects – Manus Documentation.” Manus.im,

[10] “OpenClaw proves agentic AI works. It also proves your security model doesn’t.” VentureBeat, 30 Jan. 2026,

[11] Lemos, Robert. “OpenClaw AI Runs Wild in Business Environments.” Dark Reading, 30 Jan. 2026,

[12] Vijayarangakumar, Mridula. “OpenClaw AI Agents 2026: Your New Assistant, or a Security Disaster?” Frontline, 31 Jan. 2026,

How MIT Taught AI to Read Like a Human with Recursive Language Models (RLM)

Have you ever asked an AI to analyze a long report or a big document, only to get a summary that misses the most important details? It’s a common problem. Even the most powerful AIs today can get lost when you give them too much information at once. They start to “forget” key facts buried in the text, making their answers unreliable. This has been a major roadblock, forcing us to break large documents into smaller pieces and feed them to the AI one at a time.

But what if there was a smarter way? Imagine an AI that could read a massive document like a human researcher skimming for important sections, searching for keywords, and then diving deep to find the exact information it needs. That’s the revolutionary idea behind a new AI design from MIT called a Recursive Language Model (RLM), and it’s changing what’s possible with artificial intelligence.

From Reading Everything to Smart Investigation

Most AIs today work by trying to stuff as much information as possible into their short-term memory. The more you give them, the more diluted their attention becomes, and they start making mistakes. It’s like trying to drink from a firehouse; you’re bound to miss a lot.

RLMs take a completely different approach. Instead of just reading a document from start to finish, the AI acts like a detective investigating a case. It treats the document as a crime scene to be actively explored.

Here’s a simple breakdown of how it works:

  1. The Document Becomes a Searchable Space: The entire document is made available to the AI, but it doesn’t read it all at once. It’s more like having a huge library at its disposal.
  2. The AI Becomes a Problem-Solver: The main AI gets the user’s question (e.g., “Find the total revenue in the financial report”). It then thinks about the best way to find the answer in the library.
  3. A Team of Helper AIs: The main AI can delegate smaller tasks to a team of “helper” AIs. For example, it might tell one helper to search for the word “revenue,” another to find all the tables, and a third to read the summary section. It’s like a lead detective assigning different tasks to junior detectives.
  4. Putting the Clues Together: The main AI gathers all the reports from its helpers, pieces together the clues, and comes up with a final, accurate answer.

This clever process allows the AI to focus its brainpower on the most relevant parts of the document, rather than getting bogged down by unnecessary details. This diagram shows how the main AI works with its team of helpers:

Diagram illustrating a recursive language model architecture, showing user input, large context handling with over 10 million tokens, interaction with root and sub-language models, execution in a REPL environment, and the flow leading to the final answer.

By breaking down a big problem into smaller, manageable steps, the AI can solve incredibly complex questions that would stump other systems.

Flowchart illustrating a seven-step process for querying a language model, including loading context, receiving a query, generating Python code, searching context, calling sub-models, combining results, and finalizing the answer.

The Surprising Results: Smaller, Smarter, and Better

The most amazing part of the MIT research is that this new method works incredibly well. In a head-to-head challenge, an RLM using a smaller, less powerful AI model beat a much larger, more expensive model by over 114%.

Bar chart comparing OOLONG benchmark performance scores for GPT-5 (30.2 points), GPT-5-mini (20.3 points), and RLM(GPT-5-mini) (64.7 points), highlighting a 114% improvement for RLM over GPT-5.

This shows that a smarter approach is far more effective than just building a bigger AI. The RLM’s advantage grows even larger when dealing with enormous documents. While other AIs get confused and their performance drops, the RLM stays sharp, even when searching through the equivalent of 10 million pages of text.

Bar graph comparing performance percentages of Traditional LLMs and RLM across different context sizes (tokens), highlighting the issue of 'context rot' for Traditional LLMs at 5M tokens.

In one test that required finding information across more than 1,000 separate documents, the RLM found the correct answer every single time, while other methods failed.

Bar chart comparing accuracy percentages of different models on the BrowseComp-Plus Deep Research Task, showing the accuracy of GPT-5 (truncated) at 40%, GPT-5 + BM25 at 60%, ReAct + GPT-5 + BM25 at 80%, and RLM (GPT-5) at 100%.

See It for Yourself: A Fun, Hands-On Demo

To help everyone understand this technology, I built a web app that lets you see an RLM in action. The app gives the AI a classic “needle in a haystack” challenge: find a secret number hidden somewhere in a one milion of lines of text.

Screenshot of an RLM demo interface created by Lawrence Teixeira, showing input and execution log sections with configuration settings.

You can watch on the screen as the AI works through the problem, delegating tasks and narrowing down its search until it finds the hidden number. It’s a great way to see this new kind of AI thinking in real-time.

Why This Matters: More Power and Privacy for Everyone

This new approach does more than just improve performance. It gives more people access to powerful AI and helps solve some of the biggest problems with AI today.

  1. It solves the “Forgetting” Problem: The AI no longer gets lost in long documents.
  2. It Protects Your Privacy: Because this method is so efficient, it can run on your own computer. This means you can analyze sensitive financial or medical records without your data ever leaving your control.
  3. You’re in Charge: You don’t have to rely on big tech companies to use powerful AI. You can run it yourself, on your own terms.

For businesses, this is a game-changer. Imagine an AI that can review thousands of legal contracts for risks, or a programmer’s assistant that can find a single bug in millions of lines of code. These are the kinds of powerful tools that RLMs make possible.

What’s the limitation of RLM?

The main limitation of RLM is that its power comes with overhead and complexity. When the input is short and the task is simple, using the base model directly is often faster and more efficient, since RLM adds extra steps like environment interaction and recursive calls.

In its current form, RLM relies on synchronous, blocking submodel calls, which increase end-to-end latency and can slow responses. The paper also notes that system prompts are fixed and not tailored to different task types, leaving performance gains on the table.

Finally, letting the model write and execute code inside a REPL introduces real engineering challenges, especially around security isolation, safety, and predictable behavior.

In short, RLM is powerful for hard, large-scale problems, but it is heavier, slower, and more complex than standard models for simple tasks.

Read the Official Research Paper

If you want to dive deeper into the technical details behind Recursive Language Models, the MIT researchers have published their full findings in an official paper. You can read the complete research, including all the experiments and results, on arXiv:

Official Paper: Recursive Language Models – The full academic paper by Alex Zhang and the MIT CSAIL team.

Conclusion

Scaffolding to handle extremely long contexts is becoming increasingly important for LLMs, and context folding is a promising approach in this direction. We currently believe that the Recursive Language Model is the best method for context folding, due to its simplicity and, at the same time, great flexibility and extensibility. The future of AI isn’t just about raw power; it’s about intelligence, efficiency, and a new, recursive way of solving problems.

That’s it for today!

Sources

[1] Recursive Language Models (RLM): A New Paradigm for Retrieval-Augmented Language Modeling. (2026). Manus AI Internal Document.

[2] Zhang, A. (2025, October 15). Recursive Language Models. Alex L. Zhang.

[3] Kohli, V. (2026, January 8). Breaking the Context Window: How Recursive Language Models Handle Infinite Input. GetMaxim.ai.

[4] Gibbons, P. (2026, January 19). The MIT RLM: How to Build Powerful Sovereign AI at Home. Think Bigger Think Better.

From Locked PDFs to Limitless AI: The Plain Text Revolution You Can’t Ignore

In today’s world, we’re surrounded by data. From company reports and legal, intellectual property documents to academic papers and scanned invoices, a vast amount of our collective knowledge is stored in PDF files. For decades, PDFs have been the digital equivalent of a printed page, easy to share and view, but incredibly difficult to work with. This has created a massive bottleneck in the age of Artificial Intelligence (AI).

As a technology leader, you’re constantly looking for ways to leverage AI to drive business value. But what if your most valuable data is trapped in a format that AI can’t understand? This is the challenge that a new wave of technology is solving, and it all starts with a surprisingly simple solution: plain text.

The Surprising Power of Plain Text: What is Markdown?

If you’ve ever written a quick note on your computer or sent a text message, you’ve used plain text. Markdown is a plain-text markup language that uses characters you already know to add simple formatting. For example, you can create a heading by putting a # in front of a line, or make text bold by wrapping it in **asterisks**.

This might not sound revolutionary, but it’s a game-changer for AI. Unlike complex file formats like PDFs or Word documents, which are filled with hidden formatting code, Markdown is clean, simple, and easy for both humans and computers to read. It separates the meaning of your content from its appearance, which is exactly what AI needs to understand it.

Markdown
---
title: "Markdown.md in 5 minutes (with a real example)"
author: "Your Name"
date: "2026-01-11"
tags: [markdown, docs, productivity]
---

# Markdown.md in 5 minutes ✅

Markdown (`.md`) is a plain-text format that turns into nicely formatted content in places like GitHub, GitLab, docs sites, and note apps.

> Tip: Keep it readable even **without** rendering. That’s the magic.

---

## Table of contents

- [Why Markdown?](#why-markdown)
- [Formatting essentials](#formatting-essentials)
- [Lists](#lists)
- [Task list (GFM)](#task-list-gfm)
- [Links and images](#links-and-images)
- [Code blocks](#code-blocks)
- [Tables (GFM)](#tables-gfm)
- [Mini “README” section](#mini-readme-section)
- [Resources](#resources)

---

## Why Markdown?

- **Fast** to write
- **Portable** (works across tools)
- **Version-control friendly** (diffs are clean)

Use cases:
- README files
- technical docs
- meeting notes
- product specs
- blog posts

---

## Formatting essentials

This is **bold**, this is *italic*, and this is `inline code`.

This is ~~strikethrough~~ (supported on many platforms like GitHub).

### Headings

- `# H1`
- `## H2`
- `### H3`

### Blockquote

> “Markdown is where docs and code finally get along.”

### Horizontal rule

---

## Lists

### Unordered list

- Item A
- Item B
  - Nested item B1
  - Nested item B2

### Ordered list

1. Step one
2. Step two
3. Step three

---

## Task list (GFM)

- [x] Write the first draft
- [ ] Add screenshots
- [ ] Publish the post

---

## Links and images

### Link

Read more: [My project page](https://example.com)

### Image

![Alt text describing the image](https://placehold.co/1200x630/png?text=Markdown+Example)

> Tip: If your platform doesn’t allow external images, use local paths:
> `![Diagram](images/diagram.png)`

---

## Code blocks

### Python (syntax-highlighted)

```python
def summarize_markdown(text: str) -> str:
    return f"Markdown length: {len(text)} chars"

Why AI Loves Markdown: A Non-Technical Guide to Token Efficiency

To understand why AI prefers Markdown, we need to talk about something called “tokens.” You can think of tokens as the words or parts of words that an AI reads. Every piece of information you give to an AI, whether it’s a question or a document, is broken down into these tokens. The more tokens there are, the more work the AI has to do, which means more time and more cost.

This is where Markdown shines. Because it’s so simple, it uses far fewer tokens than other formats to represent the same information. This means you can give the AI more information for the same cost, or process the same information much more efficiently.

A bar graph comparing token efficiency of different file formats including JSON, XML, HTML, and Markdown, indicating that Markdown uses 30-60% fewer tokens than JSON.

As you can see, Markdown is significantly more efficient than other formats. This isn’t just a technical detail—it has real-world implications. It means you can analyze more documents, get faster results, and ultimately, build more powerful AI applications.

The “PDF Problem”: Why You Can’t Just Copy and Paste

So, why can’t we just copy text from a PDF and give it to an AI? The problem is that PDFs were designed for printing, not for data extraction. A PDF only knows where to put text and images on a page; it doesn’t understand the structure of the content.

When you try to extract text from a PDF, especially one with columns, tables, or complex layouts, you often end up with a jumbled mess. The reading order gets mixed up, tables become gibberish, and important context is lost. For an AI, this is like trying to read a book that’s been torn apart and shuffled randomly.

Side-by-side comparison of an original PDF monthly financial report and its traditional OCR output, highlighting errors in the OCR extraction process.

This is the “PDF problem” in a nutshell. The valuable information is there, but it’s locked away in a format that’s hostile to AI.

The Solution: How Modern AI Unlocks Your PDFs

Fortunately, a new generation of AI, called Vision Language Models (VLMs), is here to solve this problem. These models can see a document just like a human does. They can understand the layout, recognize tables and headings, and transcribe the content into a clean, structured format like Markdown.

This is where a tool like MarkPDFDown comes in. It uses these powerful VLMs to convert your PDFs and images into AI-ready Markdown, unlocking the knowledge within them.

Flowchart illustrating the process of converting a PDF document into Markdown using Vision Language Models (VLM). The diagram includes icons representing a PDF, images, a VLM, and Markdown.

Introducing MarkPDFDown: Your Bridge from PDF to AI

MarkPDFDown is a powerful yet simple tool that makes it easy to convert your documents into Markdown. It’s designed for anyone who wants to make their information accessible to AI, without needing a team of data scientists.

User interface of MarkPDFDown tool displaying options to convert PDF files and images into Markdown format.
MarkPDFDown – PDF/Image to Markdown Converter

With MarkPDFDown, you can:

  • Convert PDFs and images to Markdown: Unlock the data in your scanned documents, reports, and other files.
  • Preserve formatting: Keep your headings, lists, tables, and other important structures intact.
  • Process documents in batches: Convert multiple files at once to save time.
  • Choose your AI model: Select from a range of powerful AI models to get the best results for your documents.

The Script Behind the Magic

To give you a peek behind the curtain, here is a snippet of the Python code that powers MarkPDFDown. This script handles file conversion, using the powerful LiteLLM library to interface with various AI models.

Python
import streamlit as st
import os
from PIL import Image
import zipfile
from io import BytesIO
import base64
import time
from litellm import completion

# --- Helper Functions ---

def get_file_extension(file_name):
    return os.path.splitext(file_name)[1].lower()

def is_pdf(file_extension):
    return file_extension == ".pdf"

def is_image(file_extension):
    return file_extension in [".png", ".jpg", ".jpeg", ".bmp", ".gif"]

# ... (rest of the script)

This script is a great example of how modern AI tools are built—by combining powerful open-source libraries with the latest AI models to create simple, effective solutions to complex problems.

The Future is Plain Text

The shift from complex, proprietary formats to simple, plain text is more than just a technical trend—it’s a fundamental change in how we interact with information. By making our data more accessible, we’re paving the way for a new generation of AI-powered tools that can understand our knowledge, answer our questions, and help us make better decisions.

As a leader, you don’t need to be a programmer to understand the importance of this shift. By embracing tools like MarkPDFDown and the principles of AI-ready data, you can unlock the full potential of your organization’s knowledge and stay ahead of the curve in the age of AI.

That’s it for today!

Sources

Boosting AI Performance: The Power of LLM-Friendly Content in Markdown

Why Markdown is the best format for LLMs

Improved RAG Document Processing With Markdown

MarkPDFDown GitHub Repository

Lawrence Teixeira’s Blog – Tech News & Insights