Resurrecting an old blog.
2025 ended with Meta buying Manus for over $2b and Claude Code reaching a $1B run rate . Here I explore common design patterns across these and other popular agents based on many of my favorite blogs, papers, and posts. We are getting closer to long-running autonomous agents. @METR_Evals reports that agent task length doubles every 7 months . But one challenge is that models get worse as context…
Lance Martin Humans refine their skills and learn preferences through experience. But many AI agents lack this capacity for continual learning . I created a plugin called Claude Diary that gives Claude Code the ability to learn from experience and update its own memory. You can check out the code here . Agent Memory The CoALA paper by Sumers et al. (2023) proposes a framework for agent memory,…
Lance Martin Why Context Engineering Earlier this week, I had a webinar with Manus co-founder and CSO Yichao “Peak” Ji . You can see the video here , my slides here , and Peak’s slides here . Below are my notes. Anthropic defines agents as systems where LLMs direct their own processes and tool usage, maintaining control over how they accomplish tasks. In short, it’s an LLM calling tools in a loop.…
Lance Martin The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. Rich Sutton, The Bitter Lesson The Bitter Lesson has been learned over and over across many domains of AI research, such as Chess, Go, speech, vision. Leveraging computation turns out to be the most important thing…
Lance Martin TL;DR Agents need context to perform tasks. Context engineering is the art and science of filling the context window with just the right information at each step of an agent’s trajectory. In this post, I group context engineering into a few common strategies seen across many popular agents today. Context Engineering As Andrej Karpathy puts it, LLMs are like a new kind of operating…
@rlancemartin The AI Engineering World’s fair took place last week in San Francisco. After seeing a few talks live and watching a few of the recordings , I had a few takeaways about the current state of AI agents. The Rise of Ambient Agents “Ambient” agents are becoming more common. They perform tasks autonomously and asynchronously “in the background”, often without a chat interface. Scott Wu of…
@rlancemartin Summary I compared Cursor and Claude Code on a set of coding challenges, giving them access to external documentation in 4 different ways: an llm.txt file (with and without optimized descriptions), a vector database, and direct context stuffing. A well-crafted (optimized) llms.txt file paired with Claude Code outperformed all other approaches. llms.txt leverages reasoning the in the…
I recently grabbed all crime data captured by the SFPD Crime Incident Reporting system, which is available through SF open data . It has ~1.7 million records from the past 12 years with a temporal and spatial stamp on each one. I’ve been interested in digging into some GIS data, so this is awesome! SF is broken up into 10 police districts. Each record in the data is associated with a district as…
iPython notebooks for data hacking Several notebooks that collate tools/notes for data hacking in one place: CS229 (Machine Learning), an excellent Stanford course taught by Andrew Ng. An Introduction to Statistical Learning by G. James, D. Witten, T. Hastie and R. Tibshirani. scikit-learn . Seaborn . Pandas . Apache Spark . Still in development. Links to the notebooks here: Pandas and Seaborn for…
Data to develop new drugs In the late 19th century, pathogen-immune serum was first used as a treatment against infectious disease. This lead to generations of work on antibiotics, resulting in sulfa drugs by 1936 and penicillin in 1943. Nearly all of the antibiotics in use today are compounds that were discovered during the 1940s to 1960s - the golden era of antibiotic discovery - or their…