Objective 2.1 · high
Implement agents and tools
Define bounded goals, schemas, permissions, memory, tool selection, failure handling, human approval, and observable agent execution.
- Lesson
- d2-lesson
- Practice pool
- d2-questions
- Application
- aip-l04
AIP-C01 · D2 · 26%
Implementation and Integration
Implement secure model APIs, deterministic workflows, agents, tools, enterprise integrations, deployment, and production application patterns.
Provider facts checked 2026-08-03
Objective coverage
Objective 2.2 · high
Deploy GenAI applications
Package, configure, release, scale, secure, test, roll back, and isolate model-dependent application changes.
- Lesson
- d2-lesson
- Practice pool
- d2-questions
- Application
- aip-l08
Objective 2.3 · normal
Integrate enterprise systems
Connect identity, APIs, events, data systems, search, business workflows, and records with least privilege and failure isolation.
- Lesson
- d2-lesson
- Practice pool
- d2-questions
- Application
- aip-l04
Objective 2.4 · high
Use foundation model APIs
Handle invocation, streaming, quotas, retries, token limits, structured output, errors, authentication, and model-specific behavior.
- Lesson
- d2-lesson
- Practice pool
- d2-questions
- Application
- aip-l01
Objective 2.5 · normal
Apply application patterns and development tools
Choose synchronous, asynchronous, event-driven, workflow, cache, batch, and developer-tool patterns based on constraints.
- Lesson
- d2-lesson
- Practice pool
- d2-questions
- Application
- aip-l04, aip-l08
title: "Implementation and Integration" summary: "Architecture decisions for building and integrating reliable generative AI applications with AWS services and enterprise systems."
Choose orchestration from uncertainty
Use deterministic code or a workflow when the sequence and transitions are known. Use an agent only when dynamic planning or tool selection creates meaningful value and the goal, tools, permissions, time, cost, and stop conditions can be bounded.
Every tool needs a strict input/output schema, scoped workload identity, authorization independent of model text, timeout, retry policy, idempotency behavior, observability, and safe handling of untrusted results. High-impact or irreversible operations need an authoritative approval step outside the model.
Model APIs and enterprise integration
Handle authentication, model identifiers, Regions, inference configuration, token limits, context construction, streaming, structured outputs, throttling, retries, timeouts, quotas, errors, and safe degradation. Connect identity, APIs, events, queues, workflows, data systems, and records without turning the model into the system of record.
Deployment
Version code, prompts, model/configuration IDs, tool schemas, guardrails, retrieval configuration, datasets, thresholds, and infrastructure. Test in isolation, compare with a baseline, shift traffic gradually, observe quality/safety/latency/cost, and roll back the complete release unit. Changing only the prompt in production is still a production release.
Self-check
Take one “agent” design and rewrite it as a deterministic workflow. Explain which uncertainty would justify restoring agency and which controls remain mandatory either way.