Autonomous Systems

For ownersAgents that run without a person watching. The limits matter more than the features.

Agents that run without human intervention, safely.

22 lessons

Start with lesson 1

Lesson 1 of 22, ~45 minutesThe Shift from Chatbots to Long-Horizon Agents

The lessons

  1. 01The Shift from Chatbots to Long-Horizon AgentsLong-horizon agents are AI systems that work on one task for minutes or hours, taking many steps on their own, instead of answering a single question and stopping.Learn · ~45 minutes
  2. 02STaR, V-STaR, Quiet-STaR — Self-Taught ReasoningSTaR is a training method where a model writes out its reasoning, keeps only the attempts that reached the right answer, and learns from those. Later versions add a checker and push reasoning into every word.Learn · ~60 minutes
  3. 03AlphaEvolve — Evolutionary Coding AgentsAlphaEvolve pairs a coding model with an evolution-style loop: it writes many program variants, an automatic scorer tests each one, and the best survive to be changed again. It has found new math and scheduling methods this way.Learn · ~60 minutes
  4. 04Darwin Godel Machine — Open-Ended Self-Modifying AgentsThe Darwin Godel Machine is an agent that rewrites its own code, tests each new version on coding benchmarks, and keeps an archive of the variants that scored better, so it can branch from any of them later.Learn · ~60 minutes
  5. 05AI Scientist v2 — Workshop-Level Autonomous ResearchAI Scientist v2 runs a research project end to end: it picks a hypothesis, writes code, runs experiments, makes charts and drafts the paper. One of its papers passed peer review at a workshop.Learn · ~60 minutes
  6. 06Automated Alignment Research (Anthropic AAR)Automated alignment research runs teams of AI agents, each in its own sandbox, to study how to make AI behave as intended. Their shared notes are stored outside the sandboxes so the agents cannot erase their own record.Learn · ~60 minutes
  7. 07Recursive Self-Improvement — Capability vs AlignmentRecursive self-improvement is the idea of an AI that helps build better versions of itself, with each round speeding up the next. Researchers now treat it as an engineering problem, not just a thought experiment.Learn · ~60 minutes
  8. 08Bounded Self-Improvement DesignsBounded self-improvement is the set of guardrails for an AI that changes itself: rules every edit must keep, core goals it cannot modify, several scores that all must pass, and an automatic pause when results start slipping.Learn · ~60 minutes
  9. 09The Autonomous Coding Agent Landscape (2026)This lesson surveys autonomous coding agents in 2026: tools that read a codebase, make changes and run tests on their own. The setup around the model, its tools and loop, now shapes results as much as the model itself.Learn · ~45 minutes
  10. 10Permission Modes for Autonomous AgentsPermission modes are levels of freedom you give an agent, from asking before every action to approving everything automatically. Claude Code is the example, with modes such as plan, default, acceptEdits, auto and bypassPermissions.Learn · ~45 minutes
  11. 11Browser Agents and Long-Horizon Web TasksBrowser agents are AI systems that use a web browser the way a person does: clicking, typing, reading pages and filling forms, often across many sites to finish a longer task.Learn · ~45 minutes
  12. 12Long-Running Background Agents: Durable ExecutionDurable execution means an agent's progress is saved after every step, so if a server crashes, a deploy happens or a person takes days to reply, the run picks up where it stopped instead of starting over.Learn · ~60 minutes
  13. 13Action Budgets, Iteration Caps, and Cost GovernorsCost governors are limits that stop an agent from overspending: caps on tokens per request, budgets per task, daily and monthly ceilings, step limits, cheaper models for easy work, and a shutoff when spending spikes.Learn · ~60 minutes
  14. 14Kill Switches, Circuit Breakers, and Canary TokensA kill switch turns an agent off from outside its reach. A circuit breaker pauses one behavior when it spots a bad pattern. A canary token is a fake record the agent should never touch, so touching it raises an alarm.Learn · ~60 minutes
  15. 15Human-in-the-Loop: Propose-Then-CommitPropose-then-commit is a human approval pattern: the agent saves a proposed action with its reason, data sources, impact and undo plan, a person reviews it, and only then does it run, followed by a check that it worked.Learn · ~60 minutes
  16. 16Checkpoints and RollbackCheckpoints save the agent's state after each step so another worker can take over after a crash. Rollback plans say how to undo each action, and a check after every action confirms it really happened, and only once.Learn · ~60 minutes
  17. 17Constitutional AI and Rule OverridesConstitutional AI trains a model to follow a written set of principles, with the model critiquing and revising its own answers. Anthropic's newer constitution ranks safety first, then ethics, company guidelines and helpfulness.Learn · ~60 minutes
  18. 18Llama Guard and Input/Output ClassificationLlama Guard is a small model from Meta that checks what goes into and comes out of a chatbot, flagging categories of harmful content. Tools like NeMo Guardrails add rules for how conversations may flow.Learn · ~45 minutes
  19. 19Anthropic Responsible Scaling Policy v3.0Anthropic's Responsible Scaling Policy is its public rulebook for adding safeguards as its models get more capable. Version 3.0 adds standing risk reports and roadmaps, and drops an earlier commitment to pause.Learn · ~45 minutes
  20. 20OpenAI Preparedness Framework and DeepMind Frontier Safety FrameworkOpenAI's Preparedness Framework and DeepMind's Frontier Safety Framework are each lab's plan for tracking dangerous model abilities, like acting autonomously or misleading testers, and deciding which safeguards to add.Learn · ~45 minutes
  21. 21METR Time Horizons and External Capability EvaluationMETR is an independent nonprofit that tests frontier AI models. Its time horizon measure asks how long a task, as timed for a skilled human, a model can finish about half the time.Learn · ~60 minutes
  22. 22CAIS, CAISI, and Societal-Scale RiskThis lesson covers groups that study large-scale AI risk, including the Center for AI Safety and its four risk types: misuse by bad actors, races between developers, failures inside organizations, and AI systems that act against their makers.Learn · ~45 minutes

Lessons from AI Engineering from Scratch by Rohit Ghumare and contributors, MIT license. Notes for owners by Lozo. License

Next step

Rather have it built for you?

Reading is optional. Tell me what's stuck and I'll build it into the business you already run.

Tell me what's stuck
  • A real conversation, not a pitch
  • You keep your accounts
  • Nothing sends without your yes