Building agents

From a bare model API to a working agent, one mechanism at a time.

0%

1. getting started
2. halfway
3. almost there
4. complete

explanationWhat a tool call looks like on the wiretutorialBuilding your first agenttutorialSeveral tool calls in one turntutorialStopping the loop on purposeexplanationTurning good into a rubrictutorialA golden set is the agent's regression suitetutorialRecording and grading the path the agent tookexplanationHow a running agent gets attackedtutorialGuardrails in layerstutorialGating a deploy on the evaluationexplanationProtocols for tools and for other agents

tutorialWriting a tool schema the model uses correctlytutorialWhen a tool failsexplanationRebuilding the loop on an agent SDKtutorialRetrieval as a tool the agent callsexplanationWhere an agent's memory livestutorialA model as the graderexplanationWorkflows before agentstutorialReasoning before each actiontutorialA second pass that critiques the firsttutorialWriting the plan before taking the stepstutorialOrchestration in code or by a modelexplanationWhat every extra agent coststutorialCost, caching and pinning the model

tutorialApproving a risky tool callexplanationReading failures one by oneexplanationPlan everything first or react step by step

tutorialGrowing the evaluation set from traces

Review: nothing due yet

About this course

Goals

• Builds a tool-using agent loop
• Evaluates an agent's quality systematically
• Designs and orchestrates multi-agent systems
• Runs an agent in production with guardrails

Counts

28 lessons, 86 checkpoints

Builds on

• Verifying agent work
• Governance and oversight
• Connecting tools with MCP
• Grounding and memory
• What an agent is

Build an agent from a model, a loop and tools. Progress is kept in this browser; see your progress.

Lesson plan (28 lessons, 28 live)

#	Lesson	Status	Mode	Minutes	Covers	Serves	Exercise	After	Issue
Tool use
1	What a tool call looks like on the wirebuilding-agents/how-a-model-calls-a-tool	live	explanation	20 estimate	tool-use	defines-a-tool	judge		#182
2	Building your first agentbuilding-agents/agent-loop	live	tutorial	20 estimate	agent-loop	defines-a-toolimplements-the-loop	do
3	Writing a tool schema the model uses correctlybuilding-agents/tool-schemas	live	tutorial	25 estimate	tool-use	defines-a-tool	do	Building your first agent
4	When a tool failsbuilding-agents/tool-errors	live	tutorial	25 estimate	tool-use	defines-a-toolimplements-the-loop	do	Building your first agent
5	Approving a risky tool callbuilding-agents/approving-a-tool-call	live	tutorial	30 estimate	tool-use	gates-risky-tools	do	When a tool fails	#674
6	Several tool calls in one turnbuilding-agents/parallel-tool-calls	live	tutorial	25 estimate	tool-use	implements-the-loop	do	When a tool fails
The agent loop and harness
7	Stopping the loop on purposebuilding-agents/stopping-the-loop	live	tutorial	25 estimate	agent-loop	implements-the-loop	do	Building your first agent, When a tool fails
8	Rebuilding the loop on an agent SDKbuilding-agents/loop-on-an-sdk	live	explanation	25 estimate	agent-loop	uses-an-sdk	judge	Stopping the loop on purpose, Several tool calls in one turn	#196
Agentic retrieval and memory
9	Retrieval as a tool the agent callsbuilding-agents/retrieval-as-a-tool	live	tutorial	30 estimate	retrieval-memory	adds-retrieval	do	Stopping the loop on purpose	#198
10	Where an agent's memory livesbuilding-agents/agent-memory	live	explanation	25 estimate	retrieval-memory	adds-retrieval	judge	Retrieval as a tool the agent calls	#199
Evaluation and testing
11	Turning good into a rubricbuilding-agents/rubrics	live	explanation	25 estimate	evaluation	writes-a-rubric	judge	Retrieval as a tool the agent calls	#200
12	A golden set is the agent's regression suitebuilding-agents/golden-sets	live	tutorial	25 estimate	evaluation	builds-a-golden-set	do	Turning good into a rubric	#201
13	Recording and grading the path the agent tookbuilding-agents/traces-and-trajectories	live	tutorial	30 estimate	evaluation	grades-trajectories	do	A golden set is the agent's regression suite	#202
14	Reading failures one by onebuilding-agents/error-analysis	live	explanation	30 estimate	evaluation	grades-trajectoriesprioritizes-fixesbuilds-a-golden-set	judge	Recording and grading the path the agent took	#203
15	A model as the graderbuilding-agents/llm-as-judge	live	tutorial	30 estimate	evaluation	writes-a-rubricavoids-gamed-metrics	judge	Turning good into a rubric, A golden set is the agent's regression suite	#204
Design patterns
16	Workflows before agentsbuilding-agents/workflows-before-agents	live	explanation	25 estimate	patterns	picks-a-pattern	judge	Stopping the loop on purpose, Turning good into a rubric	#676
17	Reasoning before each actionbuilding-agents/react	live	tutorial	20 estimate	patterns	picks-a-pattern	do	Stopping the loop on purpose, Turning good into a rubric
18	A second pass that critiques the firstbuilding-agents/reflection	live	tutorial	25 estimate	patterns	picks-a-pattern	do	Reasoning before each action, Turning good into a rubric
19	Writing the plan before taking the stepsbuilding-agents/planning	live	tutorial	30 estimate	patterns	picks-a-pattern	do	Reasoning before each action
20	Plan everything first or react step by stepbuilding-agents/plan-or-react	live	explanation	25 estimate	patterns	picks-a-patterncomposes-patterns	judge	Writing the plan before taking the steps, A second pass that critiques the first	#208
Orchestration and multi-agent
21	Orchestration in code or by a modelbuilding-agents/orchestrating-agents	live	tutorial	30 estimate	orchestration	justifies-orchestration-costcomposes-patterns	do	Plan everything first or react step by step	#209
22	What every extra agent costsbuilding-agents/orchestration-tax	live	explanation	25 estimate	orchestration	justifies-orchestration-cost	judge	Orchestration in code or by a model, A golden set is the agent's regression suite	#210
Guardrails and production
23	How a running agent gets attackedbuilding-agents/agent-attack-surface	live	explanation	25 estimate	production	mitigates-agent-risks	judge	When a tool fails, Recording and grading the path the agent took	#211
24	Guardrails in layersbuilding-agents/guardrails-in-layers	live	tutorial	35 estimate	production	layers-defensesmitigates-agent-risks	do	How a running agent gets attacked, Stopping the loop on purpose, Recording and grading the path the agent took	#212
25	Gating a deploy on the evaluationbuilding-agents/eval-gated-deploys	live	tutorial	30 estimate	production	gates-on-evalsmanages-cost-and-rollout	do	A golden set is the agent's regression suite, A model as the grader, Guardrails in layers	#213
26	Growing the evaluation set from tracesbuilding-agents/evals-from-traces	live	tutorial	45 estimate	production	grows-the-set-from-use	do	Gating a deploy on the evaluation	#675
27	Cost, caching and pinning the modelbuilding-agents/cost-and-provider-operations	live	tutorial	30 estimate	production	manages-cost-and-rollout	do	Gating a deploy on the evaluation	#214
28	Protocols for tools and for other agentsbuilding-agents/agent-protocols	live	explanation	25 estimate	production	mitigates-agent-risks	judge	What a tool call looks like on the wire, Guardrails in layers, What every extra agent costs	#215