Free Practice Questions for LangChain Certified Agent Engineer Certification

Understand the differences between create_agent and deepagents. create_agent is a high-level API for quickly building agents with sensible defaults, while deepagents provides more granular control and customization for complex agent architectures. Know when to use each approach based on the requirements of the agent.

Middleware in the context of agents refers to components that intercept and process requests and responses, allowing for cross-cutting concerns such as logging, authentication, and input/output transformation. Understand how to implement and configure middleware to enhance agent functionality and maintainability.

Context engineering involves managing the information available to an agent over long-running interactions. This includes techniques for summarizing conversation history, selectively retaining relevant information, and structuring context to maintain performance and coherence over extended sessions.

AGENTS.md and SKILL.md are configuration files that define agent capabilities and instructions. AGENTS.md typically contains high-level agent configuration, while SKILL.md defines specific skills or tools the agent can use. Understand how to create and modify these files to customize agent behavior.

Sandboxing refers to isolating agent execution environments to ensure security and prevent unintended side effects. This includes using containers, virtual machines, or restricted permissions to limit the agent's access to system resources and external services.

Code-based evaluators use deterministic functions to assess agent outputs, while LLM-as-judge uses a language model to evaluate quality based on criteria. Understand the trade-offs: code-based evaluators are precise and fast but may miss nuanced quality, while LLM-as-judge is flexible but can be inconsistent and costly.

Online evaluators run during agent execution in production, providing real-time feedback, while offline evaluators run on logged data or test datasets after the fact. Know when to use each: online for immediate monitoring and offline for thorough analysis and regression testing.

Running experiments involves setting up controlled tests to compare agent versions or configurations. Interpreting results requires understanding statistical significance, avoiding overfitting to noise, and using metrics that align with business goals. Learn to design experiments that yield actionable insights.

Evaluator alignment ensures that the evaluation metrics correlate with human judgment or desired outcomes. This involves calibrating LLM-as-judge prompts, validating against human annotations, and adjusting thresholds to reduce false positives and negatives.

Adding examples to a dataset involves curating representative inputs and expected outputs to improve evaluator performance and coverage. This includes selecting diverse cases, labeling correctly, and maintaining dataset versioning for reproducibility.

Understand the key terms in LangSmith Deployment: a graph defines the agent's workflow, a deployment is an instance of a graph with specific configuration, a revision is a version of a deployment, an assistant is a user-facing interface to a deployment, a thread is a conversation session, and a run is a single execution of the graph.

Multi-region serving involves deploying agents across multiple geographic regions to reduce latency and improve availability. Understand how to configure deployments to serve users from the nearest region and handle data residency requirements.

Secret handling refers to securely managing API keys, tokens, and other sensitive credentials. Best practices include using environment variables, secret management services, and avoiding hardcoding secrets in code or configuration files.

Authentication verifies the identity of a user or service, while authorization determines what actions they are allowed to perform. Understand how to implement both in agent deployments, including OAuth, API keys, and role-based access control.

Storage durability ensures that agent state and conversation history persist across restarts and failures. This involves using reliable storage backends, implementing backup and recovery procedures, and designing for data consistency.

Reading traces involves analyzing the detailed logs of agent executions to understand behavior, identify errors, and optimize performance. Traces show the sequence of steps, inputs, outputs, and intermediate states, enabling debugging and performance tuning.

Grouping traces into threads organizes related executions, such as multiple turns in a conversation, for easier analysis. This helps in understanding user journeys, tracking context, and evaluating multi-turn agent performance.

Tracking costs involves monitoring token usage, API calls, and other resource consumption to manage expenses. User sentiment analysis gauges user satisfaction from interactions, often using feedback signals or sentiment models, to improve agent quality.

Online Evals are automated evaluations that run on live traffic to detect issues in real-time, while Insights provide aggregated analytics and trends over time. Use Online Evals for immediate alerts and Insights for strategic decision-making.

Alerting involves setting up notifications for anomalies, errors, or performance degradation. Configure thresholds and escalation policies to ensure timely response to production issues without alert fatigue.