Eduardo H. Ramirez, PhD

Agent and orchestration demos

In the following demos you can explore, through examples and step-by-step animations, the components of an agent (prompts, harness, model, etc.) as well as the main strategies for orchestration and tool use.

Language Español English

1. Anatomy of an agent

A single agent from the inside: what the model answers, what the harness executes, how observations come back and how the prompt grows in each iteration.

Anatomy · case 1

Price agent

Basic action and observation loop with a search engine and a calculator.

Anatomy · case 2

Function Calling

Order status: the model returns a JSON call and the program executes it.

Anatomy · case 3

ReAct

Duplicate charge complaint: thought, action and observation over five iterations.

2. Multi-agent orchestration patterns

Each demo shows the flow at the workflow level: which agent receives which context, what it writes to the shared state, who decides the next step and where code and people step in. Each agent has a system prompt chip: hover or click to see its hypothetical instructions and tools. The names of people, companies, reference numbers and figures are fictitious.

Routing by classification

Customer service messages

Source: Anthropic

Pattern 3

Coordinator/dispatcher

Unrecognized charge and modem failure

Source: Google ADK

Pattern 4

Handoff

Internet outage and service credit

Source: Microsoft

Pattern 5

Sectioning and fan-out/gather

Reviewing a pull request

Source: Google ADK

Pattern 6

Voting

Vulnerability detection

Source: Anthropic

Pattern 7

Dynamic selection and aggregation

Analysis of a fictitious stock

Source: Microsoft

Pattern 8

Orchestrator-workers

Support incidents by region

Source: Anthropic

Pattern 9

Hierarchical decomposition

Quarterly satisfaction report

Source: Google ADK

Pattern 10

Magentic

Responding to an SRE incident

Source: Microsoft

Pattern 11

Evaluator-optimizer and maker-checker

Validating a SQL query

Source: Google ADK

Pattern 13

Iterative refinement

Optimizing a function

Refund for a duplicate charge

Source: Google ADK and Microsoft

Trace notation conventions

Label	Meaning
`PATTERN`	Pattern illustrated by the trace
`AGENTS`	Participating agents, with their tools or their function
`SHARED STATE`	Keys where the agents write their results
`STEP`, `TURN`, `ITERATION`, `ROUND`	Unit of progress, depending on the pattern
`Receives`	Context that enters the agent at that step
`Action` / `Observation`	Call to a tool and its result
`GATE`	Programmatic check between steps
`HANDOFF`	Transfer of full control to another agent
`PAUSE` / `RESUME`	Human approval point

References

Anthropic. (2024, December 19). Building effective agents. https://www.anthropic.com/engineering/building-effective-agents

Cognition. (2025, June 12). Don't build multi-agents (W. Yan). https://cognition.com/blog/dont-build-multi-agents

Google. (2025, December 16). Developer's guide to multi-agent patterns in ADK (S. Saboo). Google Developers Blog. https://developers.googleblog.com/developers-guide-to-multi-agent-patterns-in-adk/

Microsoft. (2026, February 12). AI agent orchestration patterns (C. Kittel and C. Siemens). Azure Architecture Center. https://learn.microsoft.com/en-us/azure/architecture/ai-ml/guide/ai-agent-design-patterns

Ramirez, E. (2026). Function Calling vs. ReAct: example traces [Course material].

Yan, W. (2026, April 22). Multi-agents: What's actually working [Post on X]. https://x.com/walden_yan/status/2047054554433462360