Agentic Architecture
27%of exam
AgentsRoutingChainingParallelismSubagents
Claude Code
20%of exam
MemorySettingsPermissionsHooksCI
Prompts + Output
20%of exam
SystemExamplesSchemasJSONStops
Tools + MCP
18%of exam
SchemasToolsResourcesPromptsTransport
Context + Reliability
15%of exam
RAGCachingEvalsRetriesGuardrails
Quick Facts
- Credential
- CCA-F
- Issuer
- Anthropic
- Access
- Partner network
- Focus
- Production architecture
- Users
- Claude architects
- Core
- Claude ecosystem
- Verify
- Exam portal
Simple First
Call, retrieve, chain, route, agent.
CallRetrieveChainRouteAgent
Workflow vs Agent
Workflow
- Predictable sequence
- Easier control
Agent
- Dynamic path
- Higher autonomy
Prefer simplest control
Pattern Picker
- Task is simple→Single call
- Needs facts→RAG call
- Steps known→Prompt chain
- Input classes known→Router
- Subtasks independent→Parallel calls
- Steps unknown→Agent loop
- Quality measurable→Evaluator optimizer
- Risk high→Human review
Architecture Patterns
- Single call
- Best baseline
- RAG call
- Retrieve then answer
- Prompt chain
- Ordered model steps
- Router
- Classify then dispatch
- Parallel
- Independent subtasks
- Orchestrator
- Plan then delegate
- Evaluator
- Judge then improve
Agent Controls
- Loop limit
- Bound tool cycles
- Role scope
- Narrow responsibility
- Tool scope
- Least privilege
- State
- Track decisions
- Handoff
- Define output contract
- Escalation
- Ask human
- Trace
- Log key steps
Claude Code
- CLAUDE.md
- Project memory
- settings.json
- Local configuration
- Permissions
- Gate risky actions
- Slash command
- Reusable workflow
- Hook
- Run workflow action
- Subagent
- Specialized helper
- Headless
- Noninteractive run
Schema vs Prose
Schema
- Machine parsed
- Validated fields
Prose
- Human read
- Flexible wording
Parsing needs schema
Prompt Basics
- System
- Durable constraints
- Task
- Goal and inputs
- Criteria
- Success definition
- Examples
- Show desired shape
- Context
- Relevant facts
- Delimiters
- Separate content
- Refusal
- Handle unsafe asks
Structured Output
- Schema
- Constrain fields
- Enum
- Limit choices
- Required
- Prevent omissions
- Type
- Protect parsing
- Stop sequence
- End generation
- Parser
- Validate output
- Repair
- Retry invalid JSON
Tool Safety
Name, schema, scope, auth, errors.
NameSchemaScopeAuthErrors
Tool vs Resource
Tool
- Performs action
- Needs arguments
Resource
- Provides context
- Read oriented
Action needs tool
MCP Primitives
- Tool
- Callable action
- Resource
- Readable context
- Prompt
- Reusable template
- Server
- Expose capabilities
- stdio
- Local subprocess
- HTTP
- Remote service
- Auth
- Protect access
stdio vs HTTP
stdio
- Local process
- Developer workflow
HTTP
- Remote service
- Network access
Match deployment boundary
Production Ready
Eval, retry, timeout, monitor, fallback.
EvalRetryTimeoutMonitorFallback
Eval vs Monitor
Eval
- Predeployment test
- Known cases
Monitor
- Production signal
- Live outcomes
Use both
Reliability Picker
- Output must parse→Schema validation
- Facts may change→Retrieve current source
- Latency matters→Smaller model
- Cost matters→Cache prefix
- Tool may fail→Structured error
- Safety matters→Guardrails
- Behavior drifts→Regression evals
- Ambiguity remains→Ask user
Context + Retrieval
- Context window
- Token budget
- Compression
- Keep essentials
- Retrieval
- Fetch relevant facts
- Chunking
- Split documents
- Citation
- Expose provenance
- Caching
- Reuse static prefix
- Freshness
- Check current facts
Production Reliability
- Eval
- Measure behavior
- Golden set
- Known examples
- Regression
- Catch drift
- Retry
- Handle transient failure
- Timeout
- Bound latency
- Fallback
- Degrade gracefully
- Monitor
- Track outcomes
Common Traps
Overagenting
Simple call works ≠ Avoid needless loops
Broad Tools
Scope permissions ≠ Constrain inputs
Unparsed Output
Validate schemas ≠ Repair failures
Stale Context
Retrieve current facts ≠ Cite sources
No Evaluation
Define success ≠ Track regressions
Last Minute
- 1.Start with baseline
- 2.Pick pattern deliberately
- 3.Scope tool access
- 4.Validate structured output
- 5.Handle tool errors
- 6.Budget context tokens
- 7.Add eval coverage
- 8.Plan fallback behavior
Buy on Amazon
Take courses on Udemy