The Agentic Wiki

Decompose a task into a sequence of LLM calls with gate checks between steps for validation and control.

Classify incoming requests and direct them to specialized handlers for optimized, domain-specific processing.

Run multiple LLM calls simultaneously via sectioning or voting to reduce latency and improve reliability.

A central LLM dynamically decomposes tasks, delegates to worker LLMs, and synthesizes results for complex open-ended problems.

A generator-evaluator feedback loop that iteratively refines output until quality thresholds are met.

An LLM reviews its own output to identify errors and improvements through self-critique and revision cycles.

Delegate work agent-to-agent with an explicit context transfer contract.

Pause execution for review and approval at explicit checkpoints.

Generate a multi-step plan as an artifact, then execute steps with tools.

Separate generator, critic, and refiner roles and iterate until convergence.

Multiple peer agents collaborate (or argue) in a shared conversation.

Learned orchestrators that dynamically route and control agents based on task state.

Automatically discover and optimize workflow structures via search (MCTS, evolutionary).

Guidance for combining patterns without creating brittle systems.

Agents create, prioritize, and execute their own task queues.

Multi-agent networks organized as DAGs with scaling behaviors across topologies.

Agents exchange compressed latent representations instead of natural language.