Is LoopCraft free to use?

Yes. LoopCraft offers a free tier for designing and exporting AI agent loop patterns. A Pro plan is available for advanced features.

What is Loop Engineering?

Loop Engineering is designing the control flow for AI agents: how they plan actions, observe results, and iterate. LoopCraft includes 11 pre-built patterns based on research from Anthropic, OpenAI, and others — including ReAct, Plan-Execute, Evaluator-Optimizer, Multi-Agent, and more. Design loops visually and export to LangGraph, OpenAI Agents SDK, AutoGen, or CrewAI.

Which frameworks can I export to?

LoopCraft exports to LangGraph (Python), OpenAI Agents SDK (TypeScript), AutoGen (Python), and CrewAI (Python). Each export produces working code you can run directly in your project.

Do I need to write code?

No. LoopCraft is a visual builder — you design loop patterns with a drag-and-drop interface and configure nodes without writing code. Code export to LangGraph, OpenAI SDK, AutoGen, and CrewAI is generated automatically and is optional.

What loop patterns are supported?

LoopCraft supports 11 agent loop patterns: ReAct (think-act-observe cycle), Prompt Chaining (sequential steps with quality gates), Routing (classify and direct inputs), Parallel Sectioning (independent subtasks in parallel), Parallel Voting (multiple runs for confidence), Orchestrator-Workers (dynamic task decomposition), Evaluator-Optimizer (generate-evaluate loop), Plan-and-Execute (plan then execute with replanning), ReWOO (variable passing between tasks), LLM Compiler (DAG-based parallel execution, 3.6x speedup), and Multi-Agent (collaborative agents with roles).

Approval gates in agent loops: where to put a human, and how to keep the loop moving (2026-08-21)

Every team that ships agent loops reaches the same fork in the road: run everything unattended and risk a bad action reaching the outside world, or require a human everywhere and lose the point of automating. The working middle is the approval gate — one or two well-chosen pause points where a human decides, and autonomy everywhere else. Where you place those gates matters more than how many you add. Put gates before irreversible actions, not after uncertain ones. The criterion is reversibility, not difficulty. Publishing a page, sending an email, spending budget, deleting a record — these cross a boundary that no retry un-crosses, and they are exactly where a gate belongs. By contrast, a summarizer producing a mediocre draft is a quality problem, and the fix is a better exit condition or an evaluator node, not a human reading every draft. A loop with a gate on every node is not governance; it is a very expensive form. Design the review payload for a thirty-second decision. A reviewer who must reconstruct context from a raw JSON blob will either rubber-stamp or block everything. The approval node should assemble the decision surface itself: the pending action in one line, the input that produced it, the loop's own assessment of its confidence, and the exact thing being approved. If the output is a message, show the message as it will be sent. If it is a spend, show the amount and the remaining budget. Thirty seconds of real review beats thirty minutes of archaeology. Make waiting safe with timeouts and fallbacks. An unbounded pause converts an automation problem into a queue nobody watches. Every gate needs an explicit timeout policy: fail the run when the decision matters, skip the step when it is optional, or auto-approve when the action is low-risk and the delay costs more than the risk does. The policy belongs in the loop definition, decided calmly at design time, not invented by whoever notices the stuck run at midnight. Keep the gate's record as good as the loop's. Every decision — approve, reject, edit — should land in the trace with who decided, when, and what changed. That record is what makes gates auditable, and it is what lets you tune them: a gate where nobody ever edits anything may be ready for auto-approval; a gate where everything gets edited is telling you the upstream node needs fixing. LoopCraft's approval nodes, released today, carry this whole pattern — decision payload, edit-and-continue, timeout policies, and trace recording — at https://getloopcraft.com. Feedback: sapsap@qq.com.