New opportunity

Research Engineers, Agents

Distyl AI · San Francisco; New York, United States

About this role

Distyl AI seeks Research Engineers to bridge frontier AI research and production systems for enterprise customers. The role combines applied research, systems engineering, and customer-facing deployment to turn promising agent techniques into reliable, measurable systems for real operational workflows.

Responsibilities

Design, prototype, and implement agentic systems for complex enterprise workflows. Build compound architectures combining planning, tool use, retrieval, memory, evaluation, orchestration, and execution.
Investigate how agents reason, coordinate, recover from errors, and interact with external systems under real-world constraints. Run experiments and develop frameworks that measure task completion, reliability, robustness, and failure modes.
Create tools and abstractions for observing, debugging, testing, and improving agent behavior. Work with AI Researchers on new architectures and AI Engineers on production deployment.
Integrate agents with customer APIs, applications, data platforms, and operational workflows. Explain capabilities, limitations, tradeoffs, and risks to internal teams and customer stakeholders.

Qualifications

Experience building AI systems that use models, tools, retrieval, planning, memory, or multi-step execution to complete real tasks. Strong Python engineering skills and ability to debug complex, stateful systems.
Ability to reason across prompts, tools, context, evaluators, state, orchestration, and external APIs; design experiments to understand agent behavior; and use AI tools in coding, debugging, design, trace analysis, and experimentation.
Ability to translate ambiguous business workflows into agent designs, communicate clearly with stakeholders, and take ownership of systems’ reliability, safety, and usefulness in production. No minimum number of years or degree requirement is stated.

Work arrangement and compensation:

Based in San Francisco or New York under a hybrid model requiring 3+ in-office days per week, including Tuesday–Thursday.
Base salary is $150K–$250K, depending on experience, location, and level, with equity eligibility. Benefits include medical, dental, and vision coverage for employees and dependents; flexible time off; retirement and financial planning benefits; wellness and family-building benefits; and in-office lunches and snacks.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

AI agentsPythonAgentic systemsPlanningTool useRetrievalMemoryOrchestrationAI evaluationExperiment designDebuggingAPI integrationSystems engineeringStakeholder communication

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Build agents for real workflows
Design, prototype, and implement compound systems combining planning, tools, retrieval, memory, orchestration, and execution for complex enterprise tasks.
Measure behavior and failure
Run experiments and develop evaluations of task completion, reliability, robustness, and failure modes—not just demonstrations.
Make behavior observable
Create tools and abstractions for tracing, debugging, testing, and improving complex, stateful agent systems.
Move research into production
Explore architectures with AI Researchers, harden successful approaches with AI Engineers, and integrate agents with customer APIs, applications, and data platforms.
Translate workflows and trade-offs
Turn ambiguous customer workflows into agent designs and explain capabilities, limitations, risks, and trade-offs to stakeholders.

Likely challenges to think through

Prototype speed versus production reliability
A likely challenge is deciding when an experimental architecture is ready to harden rather than adding more capabilities. Prepare to explain how you would use traces, evaluations, and observed failures to make that decision.
Metrics versus operational success
Evaluation scores may miss whether an agent completes a real workflow safely and usefully. Prepare an approach to defining task-level success, testing failure modes, and checking that measured gains hold under realistic constraints.
External-system errors versus agent autonomy
Tool calls and customer APIs can leave multi-step execution in uncertain states; recovery must be weighed against the risk of an incorrect action. Prepare to discuss state tracking, error handling, and when an agent should stop or seek human input.
Ambiguous requests versus clear commitments
Stakeholders may describe outcomes without specifying workflow boundaries. Prepare to show how you would clarify the task, identify integration constraints, and communicate what the proposed agent can and cannot reliably do.

Practice your answers

Technical practice

How would you design an agent for a complex enterprise workflow?

What to coverThe role calls for compound systems that combine planning, tool use, retrieval, memory, evaluation, orchestration, and execution. Suggested preparation: explain which components a workflow needs, how they interact, and how you would determine whether the system completes the task reliably.

When should an agent use retrieval or memory?

What to coverThe posting names retrieval and memory as parts of agent architecture. Suggested preparation: distinguish fetching task-relevant information from retaining state across steps, and discuss how each choice could affect correctness and context.

How would you decide what an agent may do through a tool?

What to coverThe role involves tool use and execution in real customer environments. Suggested preparation: discuss tool selection, validating inputs and outputs, and the trade-off between useful autonomy and controlled execution.

How would you structure maintainable Python for a stateful, multistep agent?

What to coverThe posting emphasizes clean Python and debugging complex, stateful systems. Suggested preparation: describe how you would represent state, separate orchestration from individual steps, and test state transitions.

How would you investigate an agent that fails after an external-system error?

What to coverThe role explicitly investigates error recovery and interactions with external systems. Suggested preparation: walk through identifying the failed step, deciding whether to retry or stop, and preventing an attempted recovery from making the workflow less reliable.

How would you evaluate task completion and reliability?

What to coverThe role includes building frameworks that measure agent behavior, task completion, reliability, robustness, and failure modes. Suggested preparation: define measurable outcomes, choose representative tasks, and explain what failure analysis would reveal beyond an aggregate success rate.

How would you test whether a new agent architecture is better?

What to coverThe posting calls for experiments comparing architectures and behaviors. Suggested preparation: describe a controlled comparison, what you would hold constant, and how you would interpret gains alongside new failure modes or operational costs.

What would you need to observe to debug an unexpected agent decision?

What to coverThe role includes creating tools and abstractions for observing, debugging, and testing agent behavior. Suggested preparation: explain how you would inspect prompts, context, state, tool interactions, and traces to locate where behavior diverged from the intended workflow.

How would you integrate an agent into an existing customer workflow?

What to coverThe posting names customer APIs, applications, data platforms, and operational workflows as integration points. Suggested preparation: explain how you would map inputs, outputs, and dependencies, then verify that the integrated system remains useful under real-world constraints.

What would need to change before an agent prototype could be operated in production?

What to coverThe role bridges rapid prototyping and reliable, safe production systems in collaboration with AI Engineers. Suggested preparation: discuss evidence from evaluations and traces, maintainability, operational failure handling, and the trade-offs you would surface before deployment.

Behavioural practice

Tell me about a time you turned an ambiguous workflow into a concrete system design.

What to coverThe role calls for translating ambiguous business workflows into agent designs. Suggested preparation: use a real situation, explain how you clarified the work and made design decisions, and state the resulting outcome.

Tell me about a time you worked across research and engineering to advance an idea.

What to coverThe posting describes collaboration with AI Researchers on architectures and AI Engineers on production use. Suggested preparation: describe the situation, your specific actions in coordinating or resolving a trade-off, and the result.

Tell me about a time you explained a system’s limitations or risks to stakeholders.

What to coverThe role requires clear communication with internal teams and customers about capabilities, trade-offs, and risks. Suggested preparation: give a real example showing the audience, how you made the limitation understandable, and what decision or outcome followed.

Tell me about a time you took ownership of a system reliability problem.

What to coverThe posting emphasizes responsibility for reliable, safe, useful production behavior. Suggested preparation: describe the problem, the actions you personally took to investigate and improve it, and the measured or observed result.

Tell me about a time evidence changed your view of an AI system’s behavior.

What to coverThe role favors demonstrations, traces, evaluations, and production behavior over abstract descriptions. Suggested preparation: recount a real situation, how you used evidence to revise your approach, and what changed as a result.

Sharpen your knowledge

Topics to prepare

Compound agent architecture
The posting calls for architectures combining planning, tool use, retrieval, memory, evaluation, orchestration, and execution. Revision exercise: sketch how these components would interact in a multi-step enterprise workflow, and explain where a simpler design might suffice.
State, context, and system-level behavior
The role requires reasoning across prompts, tools, context, evaluators, state, orchestration, and external APIs. Revision exercise: trace a task from initial request through successive tool results, marking how state and context change and where errors could propagate.
Agent evaluation and experiment design
The posting explicitly emphasizes experiments and frameworks measuring task completion, reliability, robustness, and failure modes. Revision exercise: define an evaluation set, success criteria, and a controlled comparison for two agent designs; explain what the results would not establish.
Error recovery and production reliability
Distyl asks engineers to investigate recovery from errors under real-world constraints and take responsibility for reliable, safe production behavior. Revision exercise: analyze an external-system failure and specify what the agent should detect, retry, stop, or surface for review.
Enterprise workflow tradeoffs
The role translates ambiguous customer workflows into agent designs and communicates capabilities, limitations, tradeoffs, and risks. Revision exercise: take a loosely specified operational task, identify questions needed to define its boundaries, and compare two possible agent designs.

Skills to revise

Maintainable Python agent implementation
Strong Python engineering and experience building systems that complete real tasks are stated qualifications. Revision exercise: implement a small multi-step agent with explicit state and tool interfaces, then add tests for a failed step.
Evaluation harness construction
The role develops evaluation frameworks and uses experiments to understand agent behavior. Revision exercise: run a fixed set of tasks through two versions of a prototype, record completion and failure categories, and inspect cases where their outcomes differ.
Trace-based debugging
The posting calls for tools and abstractions to observe, debug, and test agents, and mentions using AI tools to analyze traces. Revision exercise: capture a step-by-step execution trace, locate a failure across context, state, and tool calls, and verify a proposed fix with a repeat run.
External API integration
Integrating agents with customer APIs, applications, data platforms, and workflows is an explicit responsibility. Revision exercise: connect a prototype to a simulated API, handle an invalid response and an unavailable endpoint, and show how each outcome affects execution.
Stakeholder-facing design walkthrough
The posting expects clear communication with internal teams and customer stakeholders. Revision exercise: present a prototype's workflow, evaluation results, limitations, and risks in a short walkthrough supported by a trace rather than unsupported claims.

Your preparation checklist

Prepare an agent-system walkthrough
The role calls for systems that complete real tasks. If you have a relevant project, prepare a diagram or demo showing the workflow, your contributions, and how models, tools, retrieval, planning, or memory interacted. Label any proposed example as hypothetical.
Sketch a compound-agent design
As a practice exercise, turn an ambiguous enterprise workflow into a design with defined inputs, state, planning, tool calls, retrieval, orchestration, and execution. Note which components you would omit and why; this is preparation, not a stated interview exercise.
Build an evaluation brief
The posting emphasizes task completion, reliability, robustness, and failure modes. Prepare sample tasks, measures, a comparison between two approaches, and an example of how results would change the design.
Show a debugging trail
Prepare an available trace or clearly labeled illustrative failure case. Walk through how you would isolate a problem across prompts, context, state, tools, and external APIs, then test a Python fix; note where AI tools helped if applicable.
Plan a production handoff explanation
Outline how you would integrate an agent with a customer API or data platform, observe its behavior, recover from errors, and communicate limitations and risks. The role involves customer stakeholders, AI Researchers, and AI Engineers.
Write targeted employer questions
Ask how the team defines successful agent behavior in customer workflows, investigates production failures, and divides architecture research from deployment hardening. These questions address responsibilities in the posting without assuming Distyl’s processes.

Availability last checked: Nov 4, 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in United States. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…