BitBootcamp
☍ All concepts
Catalog
⚙ Settings
The moment a model stops answering questions and starts taking actions, running loops, calling tools, and changing real systems on its own. That autonomy is the whole payoff and the whole risk.
Part of: PAPERS-03 · Deep Reinforcement Learning, How Machines Learned to Act, AIM-05 · Reinforcement Learning, AIM-06 · Agentic AI
Interactive, in-browser playgrounds for this concept.
MuZero shows a model learning to plan and act toward a goal without being handed the rules, the intellectual root of a model that does rather than just answers.
Build up from the reward-and-planning foundations of RL to the agent's own machinery, the perceive-think-act loop, its anatomy, and the ReAct, planning, and reflection behaviors that let it reason then take action.
An agent whose loop condition is a fresh model judgment can retry 9,000 times overnight and half-write a table 9,000 times, the runaway-loop failure that mirrors the Replit agent deleting a production database.
Wrap the autonomy in external guardrails the model cannot overrule, loop and spend budgets, circuit breakers on repeated outcomes, idempotent side effects, and a defined halt-and-escalate path.
One page collapsing the agent loop, the ReAct/plan/reflect patterns, and the non-negotiable safety limits into a reference you can act from.
Every concept connects to the ideas that made it, the ways it fails, and the disciplines that keep it alive.
Browse all concepts