Teaching agents how to work without overloading their memory
Day 2 connected agents outward (tools, APIs, MCP). Day 3 looks inward: what the agent keeps in mind while working.
If you stuff every instruction into one giant prompt, the agent gets confused — wrong tools, forgotten rules, made-up details. Today you learn a cleaner pattern: skills (playbooks) and workflows (step-by-step flows).
Everything the model can “see” when it answers: system instructions, chat history, file snippets, tool results, and loaded guides. Context has a size limit (token budget).
When too much context makes performance worse — the model ignores rules, picks wrong tools, or hallucinates. More text is not always better.
Like cramming an entire textbook into your head right before one exam question — you miss the important line.
A small folder (often with a SKILL.md file) that teaches the agent how to do one job step by step — e.g. format commits, validate a database schema, run a security checklist.
Skills = recipe cards. Tools/MCP = oven and ingredients.
Show only a short summary of each skill at first; load the full instructions only when that skill is needed. Saves memory and reduces rot.
Remembering how to do something (steps), not just what happened (events) or facts (knowledge). Skills are a form of procedural memory for agents.
Code libraries that structure agents: nodes, tools, state, and flows. This course uses ADK 2.0 with graph-style workflows.
A flowchart of steps: “first classify the question → then route to FAQ or decline.” Some steps use the LLM; some use plain code for speed and reliability.
A command-line toolkit to scaffold, lint, run, test, and later deploy agent projects — not the same as Antigravity CLI (agy), which is for general coding chat. Think: “project manager for ADK agents.”
| Skill (playbook) | Tool / MCP (capability) | |
|---|---|---|
| Provides | Know-how: steps, rules, examples | Action: API call, DB query, file read |
| Simple test | If you remove the skill, the model might still do the task clumsily | If you remove the tool, the action becomes impossible |
| Examples | “How we write commit messages at our company” | “Fetch row from database” |
| Category | What it does | In this course | Same idea elsewhere |
|---|---|---|---|
| Portable skill / rule packs | Reusable instruction folders the agent loads on demand. | Agent Skills (SKILL.md) |
Cursor rules, Claude PROJECT.md / skills, Copilot instructions, custom GPT knowledge |
| Agent dev CLI | Scaffold projects, run playground, lint, eval, deploy. | Agents CLI (google-agents-cli) |
LangGraph CLI patterns, framework-specific CLIs, Makefile/scripts around your agent |
| Agent framework | Code structure for multi-step agents. | ADK 2.0 (Workflow, LlmAgent, nodes) | LangGraph, CrewAI, AutoGen, OpenAI Agents SDK |
| Local test UI | Chat with your agent while developing. | agents-cli playground | LangSmith playground, custom Streamlit/Gradio UIs, ADK web UI |
What you will get from it: Why skills spread quickly, how they fight context rot, and why one flexible agent can play many “roles” without building ten separate bots.
Main ideas: Skill folder format, progressive disclosure, builder vs developer concerns, evaluating whether skills trigger correctly, and security when importing skills from the internet.
Concept practiced: Authoring skills from simple formatters up to procedural validators; installing skill packs.
Transferable skill: Same pattern as “rules files” or “custom instructions” in any agent IDE — one markdown spec + optional scripts.
Codelab 2: Build agents with Agents CLI and ADK 2.0 codelabs.developers.google.com/agents-cli-adk-lifecycleConcept practiced: Full local dev loop — scaffold a workflow agent (classify → route → answer), lint, playground test with hot reload, CLI single-shot test.
Transferable skill: Every framework needs: project template → run locally → inspect steps → fix graph logic.
Deep dive on skills security (trust tiers, scanning shared skills), when to promote a skill to an MCP tool, and meta-skills (agents improving their own playbooks).