Home › Courses › Autonomous Agents & Frontier Safety

Autonomous Agents & Frontier Safety

Agents that run for hours, rewrite their own code, and have to be stoppable: autonomy, self-improvement, containment, and governance

What changes when an agent stops answering turns and starts running unattended. The course covers long-horizon agents, the 2026 coding-agent landscape, permission modes and durable execution; then systems that improve themselves — the STaR family, AlphaEvolve, the Darwin Gödel Machine, AI Scientist v2 — and where recursive self-improvement stops being a metaphor. The second half is the engineering that makes autonomy survivable: cost governors, kill switches and canary tokens, propose-then-commit review, checkpoints and rollback, constitutional rules and input/output classification, and the published frontier frameworks — Anthropic's RSP, OpenAI's Preparedness Framework, DeepMind's FSF and METR's time-horizon evaluations — that decide what gets deployed.

4 phases · 22 lessons · 22 worked examples · Early access

Take Agentic AI Engineering first — this course builds on it.

Every phase, every lesson, every project

• Long-Horizon Autonomy (5 lessons) — free — What changes when an agent runs for hours instead of turns: the shift from chatbots to long-horizon work, the 2026 coding-agent landscape, permission modes, browser agents, and durable execution.
  • The Shift from Chatbots to Long-Horizon Agents
  • The Autonomous Coding Agent Landscape (2026)
  • Permission Modes for Autonomous Agents
  • Browser Agents and Long-Horizon Web Tasks
  • Long-Running Background Agents: Durable Execution
• Self-Improving Systems (7 lessons) — Agents that rewrite their own reasoning or their own code: the STaR family, AlphaEvolve, the Darwin Gödel Machine, AI Scientist v2, automated alignment research, and where recursive self-improvement stops being a metaphor.
  • STaR, V-STaR, Quiet-STaR — Self-Taught Reasoning
  • AlphaEvolve — Evolutionary Coding Agents
  • Darwin Godel Machine — Open-Ended Self-Modifying Agents
  • AI Scientist v2 — Workshop-Level Autonomous Research
  • Automated Alignment Research (Anthropic AAR)
  • Recursive Self-Improvement — Capability vs Alignment
  • Bounded Self-Improvement Designs
• Control & Containment (6 lessons) — The engineering that makes autonomy survivable: action budgets and cost governors, kill switches, circuit breakers and canary tokens, propose-then-commit review, checkpoints and rollback, constitutional rules, and input/output classification.
  • Action Budgets, Iteration Caps, and Cost Governors
  • Kill Switches, Circuit Breakers, and Canary Tokens
  • Human-in-the-Loop: Propose-Then-Commit
  • Checkpoints and Rollback
  • Constitutional AI and Rule Overrides
  • Llama Guard and Input/Output Classification
• Frontier Governance (4 lessons) — The published frameworks that decide what gets deployed: Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework, DeepMind's Frontier Safety Framework, METR's time-horizon evaluations, and societal-scale risk assessment.
  • Anthropic Responsible Scaling Policy v3.0
  • OpenAI Preparedness Framework and DeepMind Frontier Safety Framework
  • METR Time Horizons and External Capability Evaluation
  • CAIS, CAISI, and Societal-Scale Risk

The technologies you will use

Claude Code · Llama Guard · Temporal

Common questions

Does this course have graded labs?

Not yet. Each lesson ships the worked example it refers to and states its exercises, but nothing here is checked by the CLI the way the from-scratch courses are. The course is marked early access for that reason.

Key terms in this course

Agent

Continue your learning path

Agentic AI Engineering · Responsible AI Engineering · Multi-Agent Systems & Swarms

Start the Autonomous Agents & Frontier Safety course

Create a free account — the opening phases of 29 of 35 courses are free, no credit card. Or see Pro pricing.

All courses · Pricing · About · FAQ · Glossary