agents-autogptlisted

# Goal-Driven Autonomous Loops The autonomous-loop pattern: give an agent a goal, and let it plan, act, observe, and replan in a loop until the goal is met or it gives up. It's the purest form of "agent" — and the best teacher of why autonomy needs guardrails. ## Overview The loop has four phases: plan (decompose the goal into steps), act (execute tools), observe (read results), and critique (evaluate progress and adjust). Memory holds the plan and findings across iterations. The loop terminates on goal completion, a step budget, or a give-up condition. This pattern shines for open-ended research and exploration — and fails instructively at long-horizon reliability, which is exactly what studying it teaches. ## When to use - Open-ended research tasks: "investigate X and summarize findings." - Exploring the capabilities and failure modes of autonomous agents. - Prototyping automation for tasks too ill-defined for a fixed pipeline. - Learning exercises in agent design — the loop makes every design decision visible. ## Core concepts - **Goal representation**: the objective stated as checkable outcomes. Vague goals produce wandering loops; concrete goals produce convergent ones. - **Plan-act-observe-critique**: the core cycle. The critique step — "did that work? what did I learn?" — is what separates this from blind execution. - **Working memory**: the evolving plan, collected findings, and action history. Without it, the agent repeats itself. - **Termination cond