Amazon AIP-C01: Bedrock Models, Agents and Guardrails
Amazon Bedrock is AWS’s managed foundation-model platform for building generative-AI applications with model access, retrieval, evaluation, safety controls, and agentic capabilities. The AIP-C01 exam is the closest current AWS certification target for professionals who need production depth around foundation-model integration and generative-AI applications.
The platform has also changed materially in 2026: Amazon Bedrock itself remains fully supported, but the older Bedrock Agents service is now Bedrock Agents Classic and stopped accepting new customers on July 30, 2026. AWS directs new agentic workloads toward Amazon Bedrock AgentCore.
Bedrock provides a broad model catalog from AWS and third-party providers, and Bedrock Marketplace extends that catalog further. Model selection should be based on task quality, latency, context needs, modalities, region availability, throughput, cost, safety behavior, and licensing or data requirements. Do not choose a model only from leaderboard reputation. Evaluate it on representative business prompts and data, then record why the selected model satisfies the quality floor. AWS also provides model-evaluation capabilities so teams can compare outputs systematically rather than relying on anecdotal preference.
Model routing can be better than one universal model. Lightweight tasks may use a faster model while complex planning or coding routes to a stronger model, provided the routing decision itself is measured. Keep fallback behavior explicit when a model is unavailable, rate-limited, or removed from a region. Reliability is part of model selection, not a separate concern after launch. Region support can be decisive when data residency, latency, or disaster recovery requirements limit where a particular model may be used. Model access and provider terms also belong in the decision record so procurement or governance changes do not surprise the application team later.
Amazon Bedrock Knowledge Bases support retrieval-augmented generation by retrieving information from enterprise data sources and supplying relevant context to the model. Retrieval quality depends on data freshness, chunking, metadata, vector or hybrid search behavior, access controls, and the quality of the source documents. A model cannot compensate reliably for a knowledge base that retrieves the wrong document or exposes data the user should not see. Treat RAG as a data and authorization system as well as a prompt-engineering pattern. Metadata filtering is important for enterprise retrieval because semantic similarity should not override tenant, region, date, product, or authorization constraints.
Evaluate retrieval separately from generation. A poor answer can originate because the wrong document was retrieved even when the model summarized it correctly. Chunk size and document structure should be evaluated on the actual queries users ask. A retrieval pattern that works for short FAQs may perform poorly on long policy or technical documents. Keep citations or source references available when the application needs users or auditors to verify where an answer came from.
AWS renamed Amazon Bedrock Agents to Amazon Bedrock Agents Classic and placed the service in maintenance mode for new customers on July 30, 2026. Existing customers can continue using Agents Classic, but the model catalog in that orchestration layer is frozen from the maintenance-mode date. This status matters for current technical content because older tutorials that say ‘Bedrock Agents’ without qualification can imply the service is still the recommended starting point. Current architecture should distinguish Bedrock’s model, Knowledge Base, and Guardrail services from the legacy Agents Classic orchestration layer.
Existing customers should plan migration deliberately rather than assume Classic will gain every new model or orchestration feature. AWS states that new models released after maintenance mode will be available through AgentCore, not added to the frozen Classic catalog. This difference can become a technical blocker when an application needs a newer model capability.
AWS recommends Amazon Bedrock AgentCore for capabilities similar to Agents Classic and for new production agentic applications. AgentCore provides managed infrastructure around agent compute, memory, identity, observability, tools, gateways, and deployment while allowing either managed or code-defined agent patterns. The design decision is how to divide orchestration ownership from AgentCore tool authorization and state management: which identity invokes each tool, how the business API validates writes and how an uncertain result is reconciled. This is a production-platform decision, not merely a new name for the old service. AgentCore’s managed harness can handle orchestration, tools, memory, identity, observability, and guardrail enforcement while code-defined agents retain greater control over custom logic. Choose the managed path when standard agent behavior is sufficient; choose code-defined orchestration when the application genuinely requires custom loops, multi-agent patterns, or framework-specific behavior.
Bedrock Guardrails can evaluate user inputs and model responses using content filters, denied topics, sensitive-information filters, word filters, grounding checks, and other safeguards. Guardrails can be applied to model inference, Knowledge Bases, flows, and supported agent patterns rather than being tied to one model. A guardrail should be tested against realistic prompts because safety configuration is probabilistic and must be validated for the application.
Guardrails should be layered with deterministic application validation. A policy filter can block harmful content while a separate authorization check decides whether the user may read a record or execute a tool. Test both false positives and false negatives. A guardrail that blocks legitimate business use too often can be bypassed operationally, while a permissive configuration can fail to protect the intended risk.
Bedrock evaluations support model and knowledge-base assessment, including automatic metrics, LLM-as-judge approaches, and human evaluation workflows. Use a representative dataset that includes normal tasks, edge cases, unsafe requests, ambiguous prompts, and important business scenarios. The value is comparison: model A may be more accurate, model B may be faster or cheaper, and a retrieval change may improve grounding while reducing recall. Production decisions should be based on measured tradeoffs rather than demo quality. Version the evaluation dataset so teams know whether a score changed because the model improved or because the test set changed. Human evaluation is especially valuable for nuanced tone, usefulness, or high-stakes domain judgment where automated metrics alone may miss important failure modes.
Bedrock applications need least-privilege IAM, encryption, logging, network controls where appropriate, and careful handling of prompts, retrieved data, tool permissions, and model outputs. Agentic systems raise the consequence because tools can take actions rather than only generate text. Use scoped identities, explicit tool contracts, confirmation or human approval for high-impact actions, and traceability from user request to tool execution. Guardrails complement authorization; they do not replace deterministic permission checks. Tool execution should use the narrowest service role required for that action. An agent that can read a support case should not inherit permission to delete unrelated records. CloudTrail and application traces should make it possible to attribute important actions to the user request, agent version, and tool invocation that caused them.
The AIF-C01 exam is the foundational AWS AI path, while AIP-C01 represents professional generative-AI development. Bedrock concepts can appear at both levels but with different expectations: foundation candidates need service and use-case literacy; professional developers need implementation, testing, safety, operations, and troubleshooting depth. Choose the certification from the responsibility you want to own rather than from the number of Bedrock features you recognize. AIF-C01 candidates need to understand what Bedrock features do and where they fit. AIP-C01 candidates need to design, test, secure, deploy, and troubleshoot those components in an application. That progression mirrors the shift from service literacy to production ownership.
Current AWS documentation should control architecture decisions because the July 2026 Agents Classic maintenance change makes older tutorials status-sensitive. For new builds, think in layers: model, retrieval, tools, agent runtime, safety, identity, observability, and evaluation. That layered model remains useful even as AWS evolves the specific managed services that implement each capability. Keep older Bedrock Agents tutorials labeled as Classic if they describe that orchestration service. The model, Knowledge Base, evaluation, and Guardrail services remain current, but the recommended agent runtime path has changed.
That distinction prevents architectural guidance from becoming stale even when individual API examples still work for existing customers. Build one reference application with a model, knowledge base, guardrail, evaluation dataset, and an AgentCore-based tool workflow. Then trace how user identity and permissions propagate through the design. That integrated exercise reveals where generative-AI quality, cloud security, data governance, and agent operations intersect. Keep a clear separation between generative safety and business authorization. A guardrail can filter content, while IAM and application logic decide which data and actions the user is allowed to access. Production Bedrock design is strongest when those controls reinforce one another rather than relying on one layer to solve every risk.