MAS.S62 AI4AI: How to AI an AI for (Almost) Anything (MIT, Fall 2026)
by MIT Multisensory Intelligence Group (Paul Liang, Ao Qu)
A graduate reading course on AI systems that improve themselves: prompt optimization, harness evolution, RL and self-play, and generated environments.
Overview
MAS.S62 AI4AI, subtitled How to AI an AI for (Almost) Anything and also called Self-Evolving AI, is a Fall 2026 course at the MIT Media Lab taught by Paul Liang and Ao Qu of the MIT Multisensory Intelligence Group. It meets Wednesdays from 2:00 to 4:00pm in E15-359 across 14 weekly sessions from September 9 to December 9, and its course site is a public GitHub Pages repo that also hosts the group's Spring 2025 course. The syllabus is organized around four things that can evolve: context (prompt optimization and context management), the harness (memory, tools, skills, workflows and multi-agent systems), the model (reinforcement learning, self-play, distillation, test-time training and continual learning) and the task (synthetic tasks, simulations, evaluations and environments). Each week pairs lecture slides with a reading list. The context weeks assign DSPy, TextGrad, Reflexion, GEPA, OPRO, FunSearch, AlphaEvolve and The AI Scientist. The harness weeks assign Darwin Godel Machine, Recursive Language Models, MEM1, Meta-Harness, Toolformer, ToolLLM, Voyager, SkillsBench and Why Do Multi-Agent LLM Systems Fail?, plus a harness engineering tutorial deck. Week 2 adds a Colab tutorial. Later weeks cover self-play, environment generation, world models, evolving rewards, AI for science, multimodal agents and safe evolution. Students run a research project with proposal, midterm and final presentations. Kimi AI and Prime Intellect sponsored compute and API credits. As of October 26, 2026, slides are posted for weeks 1 to 7 and later weeks are listed without links, so self-learners get the reading list ahead of the slides.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 14 weekly sessions (Sep 9 to Dec 9, 2026); public slides and readings, plan on 3-5h per week to follow along
- Provider
- MIT Multisensory Intelligence Group (Paul Liang, Ao Qu)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓How prompt optimizers like DSPy, TextGrad, GEPA and OPRO treat an LLM's context as something you can search and improve
- ✓How LLMs work as optimizers and discovery engines in FunSearch, AlphaEvolve and The AI Scientist
- ✓Which parts of an agent harness (memory, tools, completion checks, context management) can be evolved automatically, and how
- ✓How tool learning, agent skills and multi-agent systems evolve, and why multi-agent LLM systems fail in practice
- ✓Where reinforcement learning, distillation and test-time training fit in making a model improve from its own experience
- ✓How synthetic tasks, generated environments, world models and evolving rewards feed the self-improvement loop
- ✓How to scope and present a research project on self-evolving AI through proposal, midterm and final milestones
Highlights
- •One of the first university courses organized entirely around self-evolving AI, split into context, harness, model and task evolution
- •Each week's reading list is current through 2026 and includes harness-engineering papers such as Meta-Harness, HarnessTax and Continual Harness
- •The course site lives in the public MIT-MI/how2ai-course GitHub repo (MIT license, ~500 stars), which also hosts the Spring 2025 MAS.S60 course
- •Slides for weeks 1 to 7 and a Colab tutorial are public, and the course offers a harness engineering tutorial deck in week 6
- •Compute and API credits came from Kimi AI and Prime Intellect, which signals the hands-on agent and RL work students do in their projects
Who It’s For
Best For
- ✓AI engineers who build agents and want the research behind automatic prompt, harness and memory optimization
- ✓ML researchers and grad students choosing a project in self-improving agents or agentic RL
- ✓Technical leads who need a structured reading list on self-evolving AI rather than scattered papers
Prerequisites
- •Working knowledge of LLMs, prompting and building tool-using agents
- •Comfort reading ML research papers; RL and fine-tuning basics help for the model-evolution weeks
- •Python, for the Colab tutorial and any project work
FAQ
What is MAS.S62 AI4AI: How to AI an AI for (Almost) Anything (MIT, Fall 2026)?
MAS.S62 AI4AI is MIT's Fall 2026 graduate course on self-evolving AI, the systems that keep improving after training through feedback, memory and exploration. It suits AI engineers and researchers who already build agents and want the research map behind prompt optimizers, harness search, agent memory, RL post-training and synthetic environments, with public slides and a curated paper list for each week.
Is MAS.S62 AI4AI: How to AI an AI for (Almost) Anything (MIT, Fall 2026) free?
MAS.S62 AI4AI: How to AI an AI for (Almost) Anything (MIT, Fall 2026) is free to access.
What level is MAS.S62 AI4AI: How to AI an AI for (Almost) Anything (MIT, Fall 2026) for?
MAS.S62 AI4AI: How to AI an AI for (Almost) Anything (MIT, Fall 2026) is aimed at a advanced audience. Recommended background: Working knowledge of LLMs, prompting and building tool-using agents, Comfort reading ML research papers; RL and fine-tuning basics help for the model-evolution weeks, Python, for the Colab tutorial and any project work.
How long does MAS.S62 AI4AI: How to AI an AI for (Almost) Anything (MIT, Fall 2026) take?
Expect roughly 14 weekly sessions (Sep 9 to Dec 9, 2026); public slides and readings, plan on 3-5h per week to follow along. Most learners work through it at their own pace.
What will I learn from MAS.S62 AI4AI: How to AI an AI for (Almost) Anything (MIT, Fall 2026)?
You'll learn: How prompt optimizers like DSPy, TextGrad, GEPA and OPRO treat an LLM's context as something you can search and improve; How LLMs work as optimizers and discovery engines in FunSearch, AlphaEvolve and The AI Scientist; Which parts of an agent harness (memory, tools, completion checks, context management) can be evolved automatically, and how; How tool learning, agent skills and multi-agent systems evolve, and why multi-agent LLM systems fail in practice; Where reinforcement learning, distillation and test-time training fit in making a model improve from its own experience; How synthetic tasks, generated environments, world models and evolving rewards feed the self-improvement loop; How to scope and present a research project on self-evolving AI through proposal, midterm and final milestones.
Topics
Sources
This page was written from 3 sources, 1 on domains other than mit-mi.github.io.