AI agents explained: Build your first agent in 8 minutes

Follow the codelab → https://goo.gle/3Q5TSt3 GitHub repo → https://goo.gle/4fsahT8 Google Agent Development Kit (ADK) → https://goo.gle/3Q3enqf

At the simplest level, an AI agent doesn’t just answer—it decides and takes action. In this video, Smitha goes beyond basic chatbots and demonstrates how to build a fully autonomous, self-correcting multi-agent system from scratch using Google’s Google Agent Development Kit (ADK). First, Smitha breaks down the theory behind modern agents: the ReAct Framework (reasoning and acting) and the 3 main agent patterns (sequential, reactive, and planning). Then, she jumps straight into Python to build a practical Blog Writing Agent. Watch along and learn how to combine Planner and Writer agents with validation checkers and loop agents to create an AI that catches its own mistakes and automatically retries until it gets it right.

— Google Cloud Tech YouTube channel description

An ~8½-minute developer tutorial from the Google Cloud Tech channel (presenter named in the video as Smitha, 10 June 2026). It is the wiki’s first source that pairs the ReAct theory with a hands-on Google ADK build — making it the practical, operational bookend to the 2022 ReAct paper ingested in the same batch. Where the paper proves the reason-act loop, this video shows a 2026 framework wiring it up in Python.

TL;DR

• Defines an agent as ReAct in practice. “An agent is software that doesn’t just answer, it can decide and take action.” The clearest explanation, per the video, “comes from the research paper ReAct: Synergizing Reasoning and Acting in Language Models” — the cycle of reason → act (call a tool/API) → observe → adjust is “the foundation of how modern AI agents work.” This is a direct line from the primary paper to a 2026 vendor tutorial.
• Three agent patterns (a teaching taxonomy): sequential (assembly-line, step 1→2→3 — predictable but rigid); reactive (decide in the moment from current state — flexible, doesn’t plan ahead); deliberate/planning (pause, sketch a plan, then execute — for multi-step goals with dependencies, e.g. booking travel). Choice depends on the problem: simple/predictable → sequential, dynamic → reactive, multi-step with dependencies → planning.
• Worked build: a self-correcting blog-writer (multi-agent in ADK). Topic in → structured outline → full post → alternative titles & hooks. Architecture:
  • Planner agent (LLM agent) — produces a markdown outline (title, intro, 4–6 sections, conclusion); writes to shared state under output_key: blog_outline.
  • Outline validation checker (LLM agent) — reads state; answers OK if the outline is complete, else retry with the reason.
  • robust_blog_planner — a loop agent (a workflow agent in ADK) wrapping planner + checker; re-runs the planner up to 3 times until the checker passes. A built-in safety net.
  • Writer agent — turns the saved outline into a full markdown post (audience: software engineers; explain how and why; short code snippets); saves to blog_post.
  • Blog-post validation checker + robust_blog_writer loop — same retry pattern for the draft.
  • Root agent (blogger) — exposes the two robust loops as tools; on a topic it calls planner → writer, then adds 3 alternative titles + 2 tweet-length hooks. “The root agent only has access to the two tools we exposed, so its workflow is clear and controlled.”
• The auto-correction pattern is the lesson. Validation checkers + loop agents give the system a structural retry/self-repair mechanism — the tutorial-scale echo of the wiki’s harness argument that constraints and contracts around the model (not the model alone) deliver reliable agents. ADK’s LlmAgent / LoopAgent / BaseAgent (subclass for custom agents) are the framework primitives.
• Run surface: adk web launches a local UI to interact with the agent; demoed with “generate a blog on the top 10 use cases for AI agents.”

What was actually ingested

The full manual English closed-caption track (~183 transcript segments, runtime 8:29). The fetcher also returned a duplicate auto-generated (ASR) pass which was discarded in favour of the human-curated CC. Light caption cleanup applied (ADK product names, “robust blog,” “writes will be stored”). Chapter headings from YouTube macro-markers are interleaved in the raw transcript.

Linked entities and concepts

• Concepts: react-reasoning-acting, ai-agents, agent-harness, agent-development-lifecycle.
• Entity / framework: Google (ADK; Google Cloud Tech channel). Presenter “Smitha” named in body only (single-mention, no entity — per author-promotion rule).

Relationships

• uses 2022-10-06-yao-et-al-react-synergizing-reasoning-acting — the tutorial’s conceptual foundation is this paper’s reason-act loop, cited by name.
• uses 2026-04-22-cheung-ippolito-secchi-google-agents-cli — builds on Google ADK, the harness framework that source documents (the CLI’s ADK Code skill is the Python-API surface used here).
• published-by Google.