Skip to content

AI Cookbook

Each recipe on this page is a complete, runnable AI workflow. Register the definition, run it, then swap in your own models, tools, and data. The recipes are built the way you would run them in production: loops have limits, tool access is allowlisted, risky steps wait for human approval, and every run records what happened.

AI Cookbook production starter model Two recipe categories feed a durable production starter. The starter connects model and tools through a policy control and produces an inspectable outcome. Agentic Workflows the graph decides what runs LLM · MCP · agents · humans AI Agents the agent owns its own loop SDK · guardrails · memory Production starter Model / tools / agents policy · approval · limits Inspectable outcome evidence · state · media reference

Agentic Workflows

The workflow graph is the agent. A model reasons, but Conductor decides what actually executes: LLM, MCP, and agent tasks composed with SWITCH, DO_WHILE, FORK_JOIN, and HUMAN. The allowlist of possible actions lives in the definition, not in a prompt, so a model cannot widen its own blast radius.

Each of these carries the control that makes the pattern safe to run for real — a bounded loop, an enforced allowlist, an explicit refusal path, or a human gate.

Recipe Outcome Built from
RAG Agent Retrieve, grade whether the context can answer, retry, and refuse rather than answer ungrounded. DO_WHILE, LLM_SEARCH_INDEX
MCP Tool Calling Discover tools, shortlist them, and re-check the model's choice against that allowlist. LIST_MCP_TOOLS, CALL_MCP_TOOL, SWITCH
A2A Agent Orchestration Delegate to two remote A2A agents in parallel, join, and synthesize. GET_AGENT_CARD, FORK_JOIN, AGENT
HITL Workflow Draft an action, pause for a human, and send only on explicit approval. HUMAN, SWITCH, HTTP
LLM with Guardrails Fence a model call with a pattern screen, policy checks, and one bounded repair. INLINE, SWITCH, TERMINATE
Deep Research Agent Decompose a goal, fan out searches, review coverage each round, render a PDF. DO_WHILE, FORK_JOIN_DYNAMIC, GENERATE_PDF
A2A Delegation Hand a request to an agent someone else operates, over A2A. AGENT (a2a)

AI Agents

An agent owns its own reasoning loop: it decides which tool to call and when it is done. You author it with a Conductor SDK in Python, TypeScript, Java, or C#, or bring one written in LangChain or Google ADK through the Conductor bridge. Conductor supplies what the loop cannot give itself — every tool call is a durable, individually retryable task, and approval and cancellation are boundaries the agent cannot skip.

Recipe Outcome Built from
Tool calling agent Declare two tools and let the model choose between them. SDK Agent + @tool
Agent with guardrails Check the agent's own output and retry when a rule fails. RegexGuardrail, @guardrail
Multi-agent handoff A supervisor delegates to the specialist that fits. Strategy.HANDOFF
Agent with memory Recall facts across sessions by relevance, not replay. SemanticMemory
Agent with CLI tools Run real shell commands, restricted to an allowlist. cli_allowed_commands
Massively parallel agents Fan out to 100 sub-agents and synthesize the results. scatter_gather()
Conductor agent Invoke a stable deployed capability from another workflow. AGENT (conductor)
LangChain investigator Author with LangChain and invoke through the Conductor bridge. AGENT (conductor)
ADK triage Author with ADK and invoke through the Conductor bridge. AGENT (conductor)
Specialist review Collect independent reviews with durable fan-out and join. AGENT, FORK_JOIN, JOIN
Agent approval Pause a deployed agent at its durable approval boundary. AGENT, SWITCH, HUMAN
Agent cancellation Propagate parent termination to a long-running deployed agent. AGENT, FORK_JOIN, TERMINATE

Every definition leans on Conductor's defaults for retries and timeouts, so the JSON stays readable — add explicit limits where a provider quota or blast radius demands them. Keep documents, media, and long evidence out of workflow payloads; pass object-storage or Files API references instead.