Orchestration patterns

Prompt routing

One agent, multiple personas.

Limits of single agents

Context pollution, tool overload, and confused reasoning.

Module 2

Multi-agent communication

Message passing, pub/sub, shared state, and conversation protocols.

3 lessons

Why multiple agents?

Separation of concerns for LLMs.

Message passing patterns

Direct, pub/sub, and shared state approaches.

Conversation protocols and handoffs

Structured handoffs vs free-form chat.

Module 3

Orchestration topologies

Sequential, supervisor/worker, hierarchical, and swarm patterns.

5 lessons

Sequential pipelines

Agent A feeds agent B.

Supervisor/worker pattern

Orchestrator delegates to specialists.

Hierarchical orchestration

Orchestrator > sub-orchestrators > workers.

Nested orchestration loops

Supervisor loops wrapping worker loops.

Peer-to-peer and swarm patterns

Decentralized agent coordination.

Module 4

State management

Shared vs isolated state, state machines, and checkpointing.

3 lessons

Shared state vs isolated state

Tradeoffs in multi-agent state.

Modeling the orchestration loop as a state machine

PLANNING, EXECUTING, OBSERVING, REFLECTING, COMPLETE, FAILED.

Checkpointing and resumability

Making agent workflows survive restarts.

Module 5

Agent safety and control

Tool whitelisting/blacklisting, permissions, guardrails, and human-in-the-loop.

4 lessons

Tool whitelisting and blacklisting

Controlling what agents can and can't do.

Permission scopes per agent

Principle of least privilege for agents.

Human-in-the-loop approval gates

Confirm before executing destructive tools.

Input/output guardrails

Sanitizing inputs and filtering outputs.

Module 6

Metacognition

Agents that reflect on their own reasoning and adapt strategies.

3 lessons

Self-reflection patterns

Agents that evaluate their own outputs.

Strategy adaptation

Changing approach mid-task based on results.

Confidence estimation and knowing when to ask for help

When the agent should escalate vs proceed.