Overview
Domain 2 is about turning a GenAI design into running software, across five task statements. Agentic AI covers autonomous agents with memory and state, ReAct-style reasoning orchestrated by Step Functions, safeguards such as stopping conditions and circuit breakers, multi-model coordination, human review steps and tool integration through the Model Context Protocol (MCP), with Strands Agents and AWS Agent Squad named in the guide. Model deployment covers Bedrock on-demand and provisioned throughput, SageMaker AI endpoints and LLM-specific container concerns. Enterprise integration covers event-driven patterns, identity federation, Outposts and Wavelength for data location, and CI/CD with a GenAI gateway. FM API integration covers synchronous, asynchronous and streaming calls, retries and routing, and the final task covers Amplify, OpenAPI, Bedrock Data Automation, Amazon Q Developer and troubleshooting tooling.
At 26%, it is the second-largest domain and the most hands-on one.
How it's tested
Domain 2 scenarios describe a working or planned GenAI application with a specific integration problem, such as runaway agents, throttling, slow responses or legacy systems. You pick the AWS pattern that solves it with sound reliability and security, often choosing among several services that could technically work.
-
Contain the agent
An agent loops, overspends or acts beyond scope, and you choose stopping conditions, timeouts, IAM boundaries or human approval.
-
Choose the invocation pattern
Latency, volume and job length are described, and you pick synchronous, asynchronous with SQS, batch or streaming invocation.
-
Host the tool
A tool must be shared with agents, and you decide between a stateless Lambda MCP server and a container-based MCP server on ECS.
A chatbot shows nothing for twelve seconds and then prints a long answer, and users abandon it. Model latency is not the main issue, because total generation time is acceptable. Switching to the Bedrock streaming API and delivering tokens over WebSockets shows text within the first second. The clue is perceived delay rather than total time, so streaming beats a faster but weaker model.
Task statements 5
Official task statements from the Amazon Web Services outline, in outline order.
2.1 Implement agentic AI solutions and tool integrationsAgents, reasoning workflows, safeguards, multi-model coordination, human review and MCP tools.
- What it means
- This task covers building autonomous agents with memory and state (Strands Agents, AWS Agent Squad, MCP), structured reasoning such as ReAct and chain-of-thought orchestrated with Step Functions, safeguards like stopping conditions, Lambda timeouts, IAM boundaries and circuit breakers, model ensembles, human-in-the-loop approval, and tool integration through function definitions and MCP servers on Lambda or Amazon ECS.
- What to understand
- Agents call tools in a loop, so they need limits: maximum iterations, timeouts and least-privilege IAM roles for each tool. Human approval steps belong in Step Functions with callback tasks. MCP standardizes how agents discover and call tools; lightweight stateless tools fit Lambda, while complex or stateful tools fit ECS. Multi-agent frameworks route work to specialist agents.
- How it shows up
- An agent design, tool requirement or failure mode is described, and you choose the orchestration, safeguard or hosting option.
- Common confusion
- Assuming an agent stops by itself; explicit stopping conditions and timeouts are required to prevent loops and runaway cost.
2.2 Implement model deployment strategiesOn-demand, provisioned throughput, SageMaker AI endpoints and LLM-specific deployment.
- What it means
- This task covers deploying FMs to fit performance needs: Lambda calling Bedrock on demand, Bedrock provisioned throughput for steady high volume or custom models, SageMaker AI endpoints for self-hosted models, container patterns tuned for GPU memory and token throughput, and cascading from small models to larger ones for routine queries.
- What to understand
- On-demand Bedrock suits variable traffic and pay-per-token billing. Provisioned throughput reserves model units for predictable heavy load and is needed to serve some customized models. SageMaker AI endpoints give control over instance type and container for open-weight models. LLMs differ from classic ML because model loading time and GPU memory dominate cost and cold starts.
- How it shows up
- Traffic shape, model type and latency target are described, and you choose the deployment option.
- Common confusion
- Buying provisioned throughput for spiky, low-volume traffic where on-demand is cheaper.
2.3 Design and implement enterprise integration architecturesLegacy connectivity, event-driven integration, secure access, hybrid locations and CI/CD gateways.
- What it means
- This task covers bringing FM capabilities into existing enterprise environments: API-based integration with legacy systems, event-driven designs with Amazon EventBridge, microservices behind API Gateway, identity federation and role-based access to models and data, AWS Outposts and Wavelength for data location needs, and CI/CD with CodePipeline and CodeBuild plus a centralized GenAI gateway.
- What to understand
- Event-driven integration decouples slow GenAI processing from core systems. Federation lets enterprise identities reach FMs with scoped roles instead of shared keys. Outposts keeps processing near on-premises data for residency rules. A GenAI gateway centralizes authentication, quotas, logging and model routing so every team consumes models consistently, and pipelines add security scans, automated tests and rollback.
- How it shows up
- An enterprise constraint such as legacy systems, data residency or many consuming teams is described, and you choose the integration pattern.
- Common confusion
- Letting each team call Bedrock directly with its own controls instead of a shared gateway with central policy.
2.4 Implement FM API integrationsSynchronous, asynchronous and streaming calls, resilience and model routing.
- What it means
- This task covers calling FMs reliably: synchronous Bedrock API requests, asynchronous processing with SDKs and Amazon SQS, request validation in API Gateway, streaming responses through WebSockets or server-sent events, exponential backoff and rate limiting, fallback models for graceful degradation, X-Ray tracing, and static or content-based routing to specialized models.
- What to understand
- Throttling exceptions are normal under load, so SDK retries with exponential backoff and jitter are expected. Long or bulk jobs belong on a queue, not in a request that can time out. Streaming improves time to first token. Routing can be static in code, dynamic in Step Functions based on content, or metric-driven toward the cheapest model that meets quality.
- How it shows up
- An API reliability, latency or routing problem is described, and you pick the integration technique.
- Common confusion
- Retrying immediately in a tight loop; that worsens throttling, while backoff with jitter spreads retries out.
2.5 Implement application integration patterns and development toolsGenAI API interfaces, accessible UIs, business system enhancement, developer tools and troubleshooting.
- What it means
- This task covers APIs tailored to GenAI (streaming, token limits, retries), front ends and API-first design with AWS Amplify and OpenAPI, no-code workflows in Bedrock Flows, business enhancements such as CRM integration and document processing with Bedrock Data Automation, developer productivity with Amazon Q Developer, agent pattern orchestration, and troubleshooting with CloudWatch Logs Insights and X-Ray.
- What to understand
- Bedrock Data Automation extracts structured output from documents, images, audio and video without custom parsing code. Amazon Q Developer generates, refactors and explains code and helps diagnose errors. CloudWatch Logs Insights queries invocation logs to find failing prompts, while X-Ray shows where latency builds up across services.
- How it shows up
- A productivity, interface or document-processing need is described, and you choose the AWS tool that fits.
- Common confusion
- Writing a custom document-parsing pipeline when Bedrock Data Automation already produces structured output.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Model Context Protocol (MCP)
An open protocol for agents to discover and call tools and data sources.
- Remember
- Standardizes tool access across agents.
- Tested as
- Tool integration questions.
- Trap
- Writing bespoke tool adapters for every agent.
-
Strands Agents
An AWS open-source SDK for building model-driven agents with tools.
- Remember
- Named in the guide for agents and custom behaviors.
- Tested as
- Agent build questions.
- Trap
- Confusing it with a hosting service.
-
AWS Agent Squad
An open-source framework for routing requests across multiple specialized agents.
- Remember
- Used for multi-agent orchestration.
- Tested as
- Multi-agent questions.
- Trap
- Using one oversized agent for every task.
-
ReAct pattern
Reason, act with a tool, observe the result and repeat.
- Remember
- Needs iteration limits.
- Tested as
- Structured reasoning questions.
- Trap
- Letting the loop run without a stopping condition.
-
Human-in-the-loop with Step Functions
Pausing a workflow for review using callback tasks.
- Remember
- Use for high-impact or low-confidence actions.
- Tested as
- Collaborative AI questions.
- Trap
- Fully automating actions that need approval.
-
Provisioned throughput
Reserved Bedrock model units billed by time.
- Remember
- Fits steady high volume and some custom models.
- Tested as
- Deployment questions.
- Trap
- Using it for sporadic traffic.
-
Streaming responses
Returning tokens incrementally as they are generated.
- Remember
- Cuts perceived latency, not total generation time.
- Tested as
- Real-time interaction questions.
- Trap
- Expecting streaming to reduce token cost.
-
Exponential backoff with jitter
Retrying throttled calls after growing, randomized delays.
- Remember
- AWS SDKs implement it.
- Tested as
- Resilience questions.
- Trap
- Immediate tight-loop retries.
-
Asynchronous processing with SQS
Queueing FM requests for workers to process later.
- Remember
- Absorbs spikes and avoids timeouts.
- Tested as
- Bulk or long job questions.
- Trap
- Holding a client request open for long jobs.
-
GenAI gateway
A central layer for auth, quotas, logging and routing to models.
- Remember
- Gives consistent enterprise consumption.
- Tested as
- Enterprise integration questions.
- Trap
- Letting each team integrate with no shared controls.
-
Bedrock Data Automation
Managed extraction of structured insights from documents and media.
- Remember
- Removes custom parsing code.
- Tested as
- Business system enhancement questions.
- Trap
- Rebuilding extraction with custom Lambda code.
-
Lambda versus ECS MCP servers
Lambda for light stateless tools, ECS for complex or long-running ones.
- Remember
- Match the host to tool weight.
- Tested as
- Model extension questions.
- Trap
- Putting heavy stateful tools on Lambda.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Invocation patterns
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Synchronous | Request waits for the full response | Simple, short interactions | Using it for long batch jobs |
| Streaming | Tokens arrive as generated | Fast time to first token | Assuming it lowers cost |
| Asynchronous with SQS | Requests queued for workers | Absorbs spikes and long jobs | Expecting instant replies |
| Batch inference | Many prompts processed as one job | Lower cost for offline work | Using it for interactive users |
Model hosting options
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Bedrock on-demand | Pay per token, serverless | Best for variable traffic | Assuming it guarantees capacity |
| Bedrock provisioned throughput | Reserved model units | Predictable high volume | Paying for idle capacity |
| SageMaker AI endpoint | Self-managed model hosting | Full control of model and instance | Choosing it when a Bedrock model suffices |
Agent safeguards
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Stopping conditions | Max iterations or goal checks | Prevents loops | Relying on the model to stop |
| Timeouts | Time limits on tools and steps | Bounds latency and cost | Leaving defaults too long |
| IAM boundaries | Least-privilege roles per tool | Limits blast radius | One broad role for all tools |
| Circuit breakers | Stop calling a failing dependency | Prevents cascading failure | Retrying forever |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Agent will stop itself
- What candidates pick
- Trust the model to end its reasoning loop.
- Why it looks right
- Agents seem goal-directed.
- Why it's wrong
- Models can repeat tool calls indefinitely; stopping conditions, timeouts and budget limits must be enforced outside the model.
Clue to watch for An agent runs up cost or repeats the same tool call.
-
Streaming saves money
- What candidates pick
- Enable streaming to reduce token charges.
- Why it looks right
- Streaming feels more efficient.
- Why it's wrong
- Streaming changes delivery timing only; the same tokens are generated and billed.
Clue to watch for Requirement is about perceived latency, not cost.
-
Provisioned for spiky traffic
- What candidates pick
- Buy provisioned throughput for occasional bursts.
- Why it looks right
- Reserved capacity sounds reliable.
- Why it's wrong
- Provisioned throughput bills for reserved time, so low or bursty volume is cheaper on demand with retries or cross-Region inference.
Clue to watch for Traffic is unpredictable and modest.
-
Retry instantly
- What candidates pick
- Loop retries as soon as throttling occurs.
- Why it looks right
- It seems like the fastest recovery.
- Why it's wrong
- Immediate retries amplify throttling; exponential backoff with jitter, plus rate limiting, lets capacity recover.
Clue to watch for ThrottlingException under load.
-
One agent for everything
- What candidates pick
- Give a single agent every tool and instruction.
- Why it looks right
- Fewer components feel simpler.
- Why it's wrong
- Large toolsets confuse tool selection; specialist agents coordinated by a supervisor or router perform more reliably.
Clue to watch for Many unrelated tasks or tool sets.
-
Direct access per team
- What candidates pick
- Let each team call FMs with its own keys and logging.
- Why it looks right
- Teams move faster alone.
- Why it's wrong
- Without a shared gateway, quotas, auditing and policy differ across teams and compliance becomes hard to prove.
Clue to watch for Many teams consuming models in a regulated enterprise.
Answer framework
Apply these steps to a Domain 2 integration scenario on AIP-C01.
- Identify the integration problem type: agent behavior, model hosting, enterprise connectivity, API reliability or developer tooling.
- For agents, add bounded loops, timeouts, least-privilege tool roles and human approval where actions have business impact. Store session state in a durable place such as DynamoDB so a failed step can resume.
- Match invocation style to the workload: synchronous for short chats, streaming for perceived speed, SQS or batch for long or bulk jobs.
- Build resilience with SDK backoff, rate limiting, fallback models and X-Ray tracing, and route requests to the model that fits content and cost.
- Prefer managed AWS features such as Bedrock Data Automation, a shared GenAI gateway and CI/CD pipelines with rollback over custom code.
Practice question
Want more like this? The AIP-C01 practice test has questions for every domain.
Checklist
0 of 13
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
This domain rewards building, so plan lab time rather than only reading. Start with the Bedrock Converse API in an SDK you know, then add streaming and handle a throttling error with backoff. Put a queue in front of a batch job so you see why asynchronous processing avoids timeouts.
Next, build a small agent with Strands Agents that calls two tools, one exposed as an MCP server on Lambda. Add a maximum iteration count, a timeout and a scoped IAM role, then wrap a risky action in a Step Functions approval step. Read the Agent Squad documentation to understand supervisor and router patterns for multiple agents.
Then cover deployment and enterprise topics by comparison: on-demand versus provisioned throughput versus SageMaker AI endpoints, and EventBridge, API Gateway and federation for integration. Finish with developer tooling, Bedrock Data Automation and troubleshooting with Logs Insights and X-Ray. Use practice questions to check that you can pick one pattern when two look plausible, which is how this domain is usually tested.
FAQ
How much of AIP-C01 is Domain 2?
Implementation and Integration is 26% of the scored content, the second-largest domain, with five task statements focused on building and connecting GenAI systems.
Do I need to know MCP for AIP-C01?
Yes. The guide names MCP for agent-tool interactions, stateless MCP servers on Lambda, container-based MCP servers on ECS and MCP client libraries.
What are Strands Agents and AWS Agent Squad?
Strands Agents is an AWS open-source SDK for building tool-using agents. Agent Squad is an open-source framework for coordinating multiple specialized agents.
When is provisioned throughput the right answer?
When traffic is steady and high enough to keep reserved model units busy, or when a customized model requires it. Variable traffic usually fits on-demand.
How should an application handle Bedrock throttling?
Use SDK retries with exponential backoff and jitter, apply rate limits at API Gateway, queue non-urgent work and fall back to an alternate model or Region.
What is a GenAI gateway?
A central layer that every team uses to reach FMs, applying authentication, quotas, logging, guardrails and model routing consistently.
Which tools help troubleshoot FM applications?
CloudWatch Logs Insights for prompt and response analysis, X-Ray for tracing calls across services and Amazon Q Developer for recognizing error patterns.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 2 knowledge
You have covered agents and MCP tools, model deployment, enterprise integration, Bedrock API patterns and developer tooling. Test whether you can choose the right pattern under exam conditions.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the AIP-C01 AI tutor Get any concept explained your way