Specialized AI agents wrapped as callable functions (tools) that can be invoked by an orchestrator agent. Creates a manager-specialists hierarchy where the orchestrator delegates to domain-specific agents.

Tree-like structure with supervisor agents managing sub-agents at different levels, forming a multi-tier decision-making hierarchy.

Self-organizing agent teams with shared working memory where agents autonomously hand off tasks to specialists. No central controller—agents decide when to transfer control based on expertise needed.

Agents compete or bid for tasks based on their capabilities and costs, with selection determined by market-like mechanisms.

Central orchestrator coordinates multiple specialized agents through various communication patterns. The orchestrator receives requests, routes to appropriate specialists, and aggregates results.

Agents collaborate through a shared knowledge space (blackboard) where they read and write information to collectively solve problems. Any agent can contribute when it has relevant knowledge.

Agents execute in a fixed, predetermined order. Each agent's output is stored in shared state and available to subsequent agents. Deterministic—not LLM-powered at orchestration level.

Multiple agents execute concurrently on independent tasks. No automatic state sharing during execution—results aggregated after all complete.

Agents execute repeatedly in a cycle until a termination condition is met—either max_iterations reached or an agent signals completion via exit_loop().

Pre-defined task graph (DAG) executed as a single non-conversational tool. Handles dependency resolution, parallel execution where safe, and persistent state management.

Two-agent quality pattern where one agent generates output and another reviews/critiques it. Often combined with Loop pattern for iterative improvement until quality threshold met.

Open protocol standard for AI agents to discover, communicate, and collaborate across different platforms and providers. Agents expose capabilities via standardized 'agent cards'.

A dedicated agent sits between worker and orchestrator, comparing what the task asked for, what operations were executed, and what the output claims. If there is a gap, it does not reject or retry. It sends the output back to the worker with a specific amendment request, such as acknowledging scope limitations.

Automatic trip-wires that halt or dampen agent execution when anomalous patterns are detected. Enforces cooldown periods after negative signals, rate-limits actions per target, and breaks runaway loops where each action creates new observations that trigger further actions.

Agent alternates between reasoning about what to do and taking actions, creating a thought-action-observation loop.

Prompts the model to show step-by-step reasoning before arriving at a final answer.

Explores multiple reasoning paths simultaneously, backtracking and pruning like tree search.

Agent generates initial response, then critiques and refines it iteratively.

Agent first creates a high-level plan, then executes each step sequentially.

Agent can invoke external tools/APIs/functions to augment capabilities.

Multiple specialized agents work together, each with specific roles and expertise.

Agent maintains short-term and long-term memory to track context across interactions.