← American Express Interview Insights

American Express·AI Engineer·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

Technical screen for an AI Engineer role at Amex focused almost entirely on how agents extend beyond what the base LLM can do. Pretty deep dive, more systems-thinking than I expected for a phone screen.

Questions Asked (7)

Q1

How does tool use and function calling work in AI agents? Walk through schema design, how the agent routes to the right tool, and what happens when a tool call fails.

System DesignTechnical Trade-offsAPI & Integrations
Author's notes

I started with the schema side and felt pretty solid there, JSON schemas, typed parameters, descriptions that actually help the model pick correctly.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around the three explicit parts of the question: schema design, routing, and failure handling. For each part, explain the core mechanism, then connect it to a real-world trade-off or example relevant to a financial services context like American Express. Show depth by discussing both the 'how' and the 'why' behind design choices.

Pro tip: Emphasize idempotency and observability in tool calls—especially for financial transactions—because retries and audit trails are critical in production systems. Mention that you'd log every tool call with a correlation ID to trace failures across distributed components.

1. Define the tool schema

Explain how each tool is described with a JSON schema: name, description, parameters (types, required fields, enums), and expected output. Stress that clear, unambiguous descriptions are crucial for the LLM to select the right tool.

2. Route the request

Describe how the agent uses the LLM to parse the user's intent and match it to a tool based on the schema. Mention techniques like function calling APIs (e.g., OpenAI functions), semantic routing, or a router model, and how the agent decides between multiple candidate tools.

3. Execute and validate

Walk through what happens after selection: the agent extracts arguments, validates them against the schema, and invokes the tool (e.g., via API call). Highlight the importance of input validation and sandboxing before execution.

4. Handle failures gracefully

Explain failure modes: invalid arguments, API errors, timeouts, or ambiguous results. Describe strategies like retries with exponential backoff, fallback tools, clarifying questions to the user, and returning structured error messages to the LLM for self-correction.

5. Close the loop with observability

Mention logging, metrics, and tracing for each tool call to debug and improve routing. Discuss how feedback from failures can be used to refine schemas or routing logic over time.

Key Points to Mention

  • JSON Schema for tool definitions: parameters, types, descriptions, and required fields.
  • Function calling APIs (e.g., OpenAI) and how the LLM outputs a structured call.
  • Routing strategies: LLM-based selection, semantic search over tool descriptions, or a hybrid approach.
  • Error handling: retries, idempotency, fallback tools, and user clarification.
  • Observability: logging tool calls with correlation IDs, monitoring success rates, and auditing.
  • Trade-offs: latency vs. accuracy in routing, strict vs. flexible schemas, and cost of retries.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

How do plug-in style or MCP-style integrations extend an agent's capabilities, and what are the security and reliability trade-offs involved?

System DesignTechnical Trade-offsAPI & Integrations
Author's notes

This one I actually liked.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining plug-in/MCP-style integrations as a way to dynamically extend an agent's toolset without retraining, then discuss how they enable modular, scalable capabilities. Structure your answer around the benefits (flexibility, speed, specialization) and then systematically address security (authentication, authorization, data privacy) and reliability (latency, error handling, versioning) trade-offs, using concrete examples relevant to financial services.

Pro tip: Emphasize that in regulated industries like finance, security and reliability are non-negotiable; propose a defense-in-depth approach with sandboxing, strict contracts, and observability to balance innovation with risk.

1. Define and Explain the Mechanism

Briefly describe how plug-in or MCP-style integrations work: they allow an agent to call external tools or services via standardized interfaces (e.g., APIs, function calls) to perform tasks beyond its core model. Highlight that this extends capabilities without retraining the model.

2. Enumerate Capability Extensions

List key benefits: access to real-time data, specialized domain tools (e.g., fraud detection, payment processing), and the ability to compose complex workflows. Mention improved modularity, faster iteration, and scalability.

3. Analyze Security Trade-offs

Discuss risks: expanded attack surface, data leakage, unauthorized access, and injection attacks. Cover mitigations: OAuth, scoped API keys, input validation, sandboxing, and audit logging.

4. Analyze Reliability Trade-offs

Address issues like network latency, third-party downtime, version mismatches, and error propagation. Suggest strategies: retries with backoff, circuit breakers, fallback mechanisms, and contract testing.

5. Conclude with Balanced Recommendation

Summarize that while plug-ins offer significant agility, they require robust governance. Propose a framework for evaluating and monitoring integrations, emphasizing observability and compliance.

Key Points to Mention

  • Standardized interfaces (e.g., OpenAPI, MCP) enable interoperability and reduce integration effort.
  • Security risks include prompt injection, data exfiltration, and privilege escalation; mitigate with least privilege and sandboxing.
  • Reliability concerns: latency, rate limits, and dependency failures; use circuit breakers and graceful degradation.
  • Versioning and backward compatibility are critical to avoid breaking agent behavior.
  • Observability (logging, tracing, metrics) is essential for debugging and auditing integrations.
  • In finance, compliance (e.g., PCI DSS, GDPR) must be enforced for any external integration.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

When would you use code execution as a tool for an agent, and what sandboxing concerns come with it?

System DesignTechnical Trade-offs
Author's notes

Said the obvious stuff: math, data manipulation, anything deterministic where you don't want the LLM to hallucinate a number.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the scenarios where code execution adds value to an agent, such as dynamic data analysis, complex calculations, or interacting with APIs. Then, discuss the critical sandboxing concerns like isolation, resource limits, and security, emphasizing a risk-based approach. Conclude with how you would implement safeguards in a production environment, especially in a regulated industry like finance.

Pro tip: Frame your answer around the principle of least privilege and mention specific sandboxing technologies (e.g., gVisor, Firecracker) to show depth. Also, tie it back to business impact, such as preventing data leaks or ensuring compliance, which resonates with American Express.

1. Identify Use Cases

Explain when code execution is beneficial: e.g., for tasks requiring dynamic computation, data manipulation, or automation that can't be hardcoded. Give concrete examples like generating reports or solving math problems.

2. Assess Risks

Outline the potential security risks: arbitrary code execution, data exfiltration, resource exhaustion, and lateral movement. Emphasize that these risks are amplified in multi-tenant or production environments.

3. Sandboxing Strategies

Describe sandboxing techniques: process isolation (containers, VMs), resource limits (CPU, memory, time), network restrictions, and filesystem isolation. Mention tools like Docker, gVisor, or WebAssembly.

4. Monitoring and Auditing

Highlight the need for logging, monitoring, and auditing of code execution to detect anomalies and ensure compliance. This is crucial for post-incident analysis and regulatory requirements.

5. Trade-offs and Best Practices

Discuss trade-offs between security and functionality, and recommend best practices like least privilege, input validation, and using managed sandbox services. Tailor to American Express's context (e.g., PCI DSS, data privacy).

Key Points to Mention

  • Use cases: dynamic data analysis, complex calculations, API integrations, and automation of repetitive tasks.
  • Sandboxing concerns: isolation (process, network, filesystem), resource limits, and preventing escape.
  • Technologies: containers (Docker), microVMs (Firecracker), gVisor, WebAssembly, and seccomp.
  • Security principles: least privilege, defense in depth, and zero trust.
  • Compliance and monitoring: audit logs, anomaly detection, and adherence to regulations like PCI DSS.
  • Trade-offs: performance overhead vs. security, and complexity vs. maintainability.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

How does retrieval function as a capability for an AI agent, and how does it differ from just stuffing more context into the prompt?

System DesignTechnical Trade-offs
Author's notes

Pretty comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining retrieval as an external memory capability that dynamically fetches relevant information, then contrast it with static context stuffing by highlighting trade-offs in scalability, cost, and accuracy. Use a concrete example (e.g., customer support agent) to illustrate how retrieval enables up-to-date, precise responses without bloating the prompt.

Pro tip: Emphasize that retrieval is not just about fetching documents but about integrating a retriever into the agent's decision loop, allowing it to decide when and what to retrieve—this shows you understand agentic design beyond simple RAG.

1. Define Retrieval as a Capability

Explain that retrieval allows an AI agent to access external knowledge sources on demand, functioning as a scalable memory system. It enables the agent to fetch only relevant information for a given query.

2. Contrast with Context Stuffing

Describe context stuffing as pre-loading all potentially relevant information into the prompt, which is limited by token limits, cost, and latency. Highlight that retrieval avoids these constraints by fetching only what's needed.

3. Discuss Trade-offs

Compare retrieval vs. context stuffing on dimensions like scalability, freshness, accuracy, and cost. Mention that retrieval introduces complexity (e.g., indexing, retrieval errors) but offers better long-term maintainability.

4. Illustrate with a Use Case

Provide a concrete example relevant to American Express, such as a customer service agent retrieving transaction details or policy documents. Show how retrieval improves response relevance and reduces hallucinations.

5. Conclude with Best Practices

Summarize when to use retrieval (dynamic, large knowledge bases) vs. context stuffing (small, static contexts). Mention hybrid approaches like caching or fine-tuning for specific scenarios.

Key Points to Mention

  • Retrieval augments the agent with external memory, enabling access to up-to-date and domain-specific knowledge.
  • Context stuffing is limited by token limits, cost, and latency, and can lead to irrelevant information and hallucinations.
  • Retrieval improves scalability, accuracy, and cost-efficiency by fetching only relevant information on demand.
  • Retrieval introduces challenges like retrieval quality, latency, and integration complexity, requiring careful design.
  • Agentic retrieval allows the agent to decide when and what to retrieve, enhancing autonomy and efficiency.
  • Hybrid approaches (e.g., retrieval + fine-tuning) can balance trade-offs for specific use cases.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q5

Explain the planner/executor pattern in agent architectures. What are the benefits and failure modes?

System DesignAdaptability & Ambiguity
Author's notes

I like this pattern so I went a bit long.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the planner/executor pattern and its role in decomposing complex tasks into manageable steps. Then discuss the benefits in terms of modularity, scalability, and robustness, and finally outline common failure modes such as planning errors, execution drift, and coordination overhead. Use concrete examples from AI agent systems to illustrate your points.

Pro tip: Emphasize how the pattern enables separation of concerns and facilitates debugging and testing, which is crucial in high-stakes domains like finance. Also, mention that at American Express, where reliability and compliance are paramount, you would incorporate validation and fallback mechanisms to mitigate failure modes.

1. Define the pattern

Explain that the planner/executor pattern separates high-level planning from low-level execution, where a planner generates a sequence of actions and an executor carries them out.

2. Describe the architecture

Outline how the planner and executor interact, often through a shared representation of tasks, and how feedback loops can refine plans.

3. Highlight benefits

Discuss benefits such as modularity, reusability, improved scalability, and the ability to handle complex, multi-step tasks.

4. Identify failure modes

Cover failure modes like planner errors (e.g., invalid plans), executor errors (e.g., action failures), and coordination issues (e.g., misalignment between planner and executor).

5. Mitigation strategies

Suggest ways to mitigate failures, such as validation, monitoring, fallback plans, and human oversight, especially in critical applications.

Key Points to Mention

  • Separation of concerns between planning and execution
  • Modularity and reusability of components
  • Scalability to handle complex tasks
  • Common failure modes: planning errors, execution errors, coordination breakdowns
  • Mitigation techniques: validation, feedback loops, fallback mechanisms
  • Real-world examples (e.g., robotics, virtual assistants, financial systems)

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q6

How does an agent discover and select the right tools at runtime, especially when the tool set is large?

System DesignTechnical Trade-offs
Author's notes

Blanked for a second on the discovery piece.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a retrieval and ranking challenge: the agent must efficiently narrow a large tool set to a small candidate list, then select the best tool based on task context and constraints. Discuss a layered architecture with semantic search, metadata filtering, and a lightweight selector (e.g., LLM-based or learned ranker), and highlight trade-offs like latency, accuracy, and cost. Emphasize the need for fallback mechanisms and continuous evaluation to handle tool evolution.

Pro tip: Mention that at scale, you should treat tool selection as a two-stage recommender system: first retrieve top-K candidates using embeddings and filters, then re-rank with a more expensive model only on that shortlist. This balances latency and accuracy, which is critical in production systems like those at American Express.

1. Clarify the problem and constraints

Acknowledge that a large tool set makes exhaustive evaluation impractical, and that runtime constraints (latency, cost, reliability) drive the design. Ask clarifying questions about tool count, update frequency, and acceptable latency.

2. Describe the discovery mechanism

Explain how tools are indexed and made discoverable: embed tool descriptions and schemas into a vector store, maintain metadata (e.g., domain, permissions, cost), and support hybrid search (semantic + keyword).

3. Outline the selection process

Detail a two-stage approach: retrieve top-K candidates via similarity search and filters, then use an LLM or learned ranker to pick the best tool based on the task, context, and constraints. Include fallback strategies if no tool is suitable.

4. Address trade-offs and evaluation

Discuss trade-offs between retrieval recall and precision, latency vs. accuracy, and cost of LLM calls. Mention offline evaluation with labeled data and online monitoring for tool selection quality.

5. Highlight scalability and maintenance

Explain how the system handles tool additions/removals, versioning, and dynamic updates without retraining. Emphasize caching, sharding, and incremental indexing for large-scale deployments.

Key Points to Mention

  • Semantic search with embeddings for tool descriptions and schemas
  • Metadata filtering (domain, permissions, cost, latency) to prune candidates
  • Two-stage retrieval and ranking to balance latency and accuracy
  • LLM-based re-ranking or a lightweight classifier for final selection
  • Fallback mechanisms (e.g., default tool, human escalation) when confidence is low
  • Continuous evaluation and monitoring of tool selection performance

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q7

How do you decide whether a new capability should be added as an external tool versus fine-tuned or baked into the model itself?

Technical Trade-offsProduct StrategySystem Design
Author's notes

This is genuinely a hard question and I said so, which I think landed okay.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Frame the decision as a trade-off between cost, latency, control, and maintainability, grounded in the specific use case and constraints. Walk through a structured evaluation that starts with problem definition, then compares external tools vs. fine-tuning vs. in-model approaches across key dimensions. Conclude with a recommendation that often favors external tools for flexibility and cost, while reserving fine-tuning for high-volume, stable tasks where latency and integration matter.

Pro tip: Emphasize that in regulated industries like finance, external tools offer better auditability, version control, and the ability to update without retraining, which often outweighs the marginal performance gains of fine-tuning.

1. Define the capability and success criteria

Clearly specify what the new capability is, how it will be evaluated (accuracy, latency, cost), and what business or user problem it solves.

2. Assess task characteristics

Determine if the task is static or evolving, requires real-time data, involves sensitive data, and how frequently it changes. This informs whether the model needs to be updated often.

3. Evaluate trade-offs across dimensions

Compare external tools, fine-tuning, and in-model approaches on cost, latency, scalability, maintainability, compliance, and time-to-market.

4. Consider organizational and technical constraints

Factor in team expertise, existing infrastructure, vendor lock-in, and regulatory requirements (e.g., explainability, audit trails).

5. Prototype and measure

Start with a lightweight external tool integration, benchmark performance, and only invest in fine-tuning if the external approach fails to meet critical thresholds.

Key Points to Mention

  • Cost and latency trade-offs: external tools add inference overhead but avoid retraining costs; fine-tuning reduces latency but requires data and compute.
  • Data freshness and dynamic requirements: external tools can access real-time data; fine-tuned models are static snapshots.
  • Maintainability and versioning: external tools allow independent updates and A/B testing; model changes require retraining and redeployment.
  • Compliance and auditability: external tools provide clear logs and control, crucial in regulated industries like finance.
  • Performance and accuracy: fine-tuning can yield higher accuracy for narrow, stable tasks, but may overfit or degrade on edge cases.
  • Hybrid approaches: use external tools for orchestration and fine-tuned models for core tasks, leveraging strengths of both.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.