• unwind ai
  • Posts
  • Agentless AI Software Engineering

Agentless AI Software Engineering

PLUS: Self-evolving memory layer for AI agents, FLUX Finetuning API

In partnership with

Today’s top AI Highlights:

  1. A self-evolving memory layer for AI agents and apps

  2. You don't always need AI agents to fix complex software bugs

  3. LLMs learn to change their own weights on the fly

  4. Black Forest Labs releases its own Finetuning API

  5. Free opensource alternative to Google AI Studio with local LLMs

& so much more!

Read time: 3 mins

AI Tutorials

LLMs are great at generating educational content and learning roadmaps, but they struggle with complex, multi-step workflows. While you could ask an LLM to create a curriculum, then separately ask it to design exercises, then manually compile resources – this process is tedious and requires constant human coordination.

In this tutorial, we'll solve this by building an AI Teaching Agent Team. Instead of isolated tasks, our AI agents work together like a real teaching faculty: one creates comprehensive knowledge bases, another designs learning paths, a third curates resources, and a fourth develops practice materials.

The user just needs to provide a topic. Everything is automatically saved and organized in Google Docs, creating a seamless learning experience without manual overhead. We are using Phidata and Composio to build our AI agents.

We share hands-on tutorials like this 2-3 times a week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.

Don’t forget to share this newsletter on your social channels and tag Unwind AI (X, LinkedIn, Threads, Facebook) to support us!

Latest Developments

Do we really need complex AI agents and tool interactions to solve real-world software engineering problems? The current trend of building increasingly sophisticated agent architectures might be overcomplicating what could be a simpler solution, while also driving up costs and reducing reliability.

Enter AgentLess, an open-source framework that proves simpler can be better. Using a streamlined three-phase approach - localization, repair, and patch validation - it achieves a 32% success rate on SWE-bench Lite through targeted LLM prompts and embedding-based retrieval. Its effectiveness in solving real-world coding challenges has led OpenAI to adopt it as their go-to approach for showcasing GPT-4o and O1’s coding performance.

Key Highlights:

  1. Problem-Solving Pipeline - AgentLess first generates a tree-like structure of your repository, then uses embedding-based retrieval to identify suspicious files. It provides only the class and function signatures to the LLM, helping pinpoint exact lines that need modification while keeping the context focused.

  2. Precise Patch Generation - Once the bug is located, Agentless generates multiple potential fixes as simple diffs – essentially search and replace instructions. This differs from systems that try to rewrite entire code blocks and reduces the risk of introducing unintended side effects.

  3. Reproduction Tests - To ensure the fixes actually work, Agentless automatically creates tests specifically designed to reproduce the original error, in addition to running existing regression tests. This ensures that a fix addresses the issue and doesn't break existing functionality.

  4. Cost Effective - AgentLess selects the best patch through a combination of majority voting and test consistency, rather than relying on complex agent-based decision trees. This delivers better results at a fraction of the cost ($0.70 per issue) compared to agent-based solutions.

Discover 100 Game-Changing Side Hustles for 2025

In today's economy, relying on a single income stream isn't enough. Our expertly curated database gives you everything you need to launch your perfect side hustle.

  • Explore vetted opportunities requiring minimal startup costs

  • Get detailed breakdowns of required skills and time investment

  • Compare potential earnings across different industries

  • Access step-by-step launch guides for each opportunity

  • Find side hustles that match your current skills

Ready to transform your income?

Llongterm brings a straightforward plug-and-play API to add persistent memory to your AI agents and apps. It sits as a middleware layer between your app and any LLM, remembering user interactions and injecting that context back into the system messages.

The standout is that every piece of stored information is human-readable, allowing for transparency and easier debugging. If you are working with chatbots, AI agents or RAG pipelines, you may want to explore Llongterm for stateful conversations.

Key Highlights:

  1. Persistent Memory Across Sessions - Llongterm’s core function is enabling your AI to remember user interactions between sessions, meaning users won't have to repeat information. This is achieved by creating a unique "Mind" for each user, which stores their interaction history and preferences. This removes the need to track conversation history yourself.

  2. LLM Agnostic - It is designed to function with any LLM; it isn’t tied to one model which provides flexibility. Llongterm is built as a middleware layer, easily integrated into your existing app with a simple npm install llongterm command. It uses the system message field of LLMs to provide context, so it fits into your existing architecture.

  3. Transparent and Debuggable - All stored memory is fully human-readable. This means you can directly inspect how information is stored, understand how it is influencing the LLM responses, and debug more efficiently. It offers observability without hidden logic or processes.

  4. Structured Memory with Timeline - Llongterm doesn't just store data; it organizes it through self-structuring and a timeline system. This creates a dynamic taxonomy of the information received and places events in time and space, making the context retrieval process more intelligent. This structured approach means that your AI can understand the evolution of conversations and user preferences over time for more contextually aware responses.

Quick Bites

Luma Labs has released Ray2, a new video generative model with 10x the compute power of its predecessor. Ray2 produces 10-second high-quality videos with amazing realism, instruction following, and physics and motion understanding. The text-to-video feature is available now to paid subscribers via Dream Machine. Image-to-video, video-to-video, and editing capabilities will be released soon.

DeepSeek has launched its AI app, powered by DeepSeek-V3, and it's free to use. The app is available on the App Store, Google Play, and major Android markets, offering features like cross-platform sync, web search, and file upload. It's completely free with no ads or in-app purchases.

Black Forest Labs has launched the FLUX Pro Finetuning API, allowing users to customize their FLUX Pro text-to-image models with as few as 1-5 images. This new API will help users build custom models that maintain the base model's versatility, while giving more control over specific objects, styles, and brands.

Sakana AI has introduced Transformer², a new self-adaptive ML system that enables LLMs to dynamically adapt to different tasks in real-time, similar to how living organisms adapt to their environment. The system analyzes each incoming task and adjusts its neural weights accordingly.

  • Uses Singular Value Decomposition to break down the model's "brain" into independent components that can be selectively enhanced or suppressed.

  • Outperformed traditional approaches in tasks like math and coding while using significantly fewer parameters compared to methods like LoRA.

  • Trained components from one model (Llama) were successfully transferred and improved another model (Mistral).

Tools of the Trade

  1. Glama Gateway: Access and manage 100+ AI models (from providers like OpenAI, Anthropic, and Google) through a single OpenAI-compatible API. It provides features like low latency, load balancing, fallback handling, usage analytics, and consolidated billing.

  2. Kiln: Free, open-source desktop application to fine-tune LLMs, generate training data, and collaborate on datasets. It supports multiple AI providers (like OpenAI, Ollama, Groq) while keeping data private and local, and offers both a graphical interface and Python library for developers.

  3. Memobase: First user profile-based memory for AI apps. It allows you to define and manage what user information your AI captures and uses to personalize user experiences, storing this information as evolving user profiles. Integrate with your existing LLM stack with minimal code.

  4. Awesome LLM Apps: Build awesome LLM apps with RAG, AI agents, and more to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos, and automate complex work.

Hot Takes

2025 predictions:
1. Consumers get sick of hearing about AI in smartphones
2. Foldables still don't take off
3. AI capabilities keeps on commoditizing, margin erosion continues
4. AI-first consumer hardware still doesn't find PMF beyond note taking and toys
5. Apple is still a safe stock to hold, but there will be some fluctuations due to the AI story not playing out
6. US enters a Golden Age
7. India stock market dampens but then reaches ATH
8. Europe keeps fading in relevance
9. Boring businesses become exciting, and vice versa
10. Bitcoin and gold perform well ~

Carl Pei

That’s all for today! See you tomorrow with more such AI-filled content.

Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!

Unwind AI - X | LinkedIn | Threads | Facebook

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉 

Reply

or to participate.