Start with a decision you want to explain
Open a repository with a question: what would make this system useful, unsafe or too expensive for the customer I have chosen? Leave with an artifact you can discuss, such as an evaluation table, a tool permission boundary or a recovery flow. A list of framework names gives an interviewer much less to inspect.
The project descriptions below come from their maintainers' documentation, reviewed 3 November 2026. These are selected study resources, not a ranking or a claim that employers require them. The exercises are ours. You can do the reading and product reasoning without running code; installation, API charges and data access require a separate decision.
| You need to explain | Start with | Leave with |
|---|---|---|
| Why a workflow needs an agent | Microsoft or Hugging Face | A bounded task and a simpler alternative |
| Whether a score is credible | Inspect or tau-bench | A run ledger and a defensible denominator |
| What a test should accept | promptfoo | An observable outcome and failure cases |
| What an agent may do | MCP | A tool contract and permission check |
| What Stop can promise | A2A | A cancellation and recovery contract |
Microsoft AI Agents for Beginners: follow the work into production
The Microsoft curriculum includes tool use, agent design patterns, production operation and memory. Read a tool-use example alongside the production lesson. Notice which parts of the demonstration depend on a working service, a clear request or an already valid permission.
Our exercise: design an equipment-request assistant for a company with an existing purchase form. Keep the form as the comparison. Describe one place where flexible interpretation helps, and keep the final order behind the company's approval rule. Then make the catalog unavailable halfway through. Write what the employee sees and who resolves the request.
Your artifact is a one-page task boundary: supported requests, external actions, approval owner and the failure that sends work back to a person. If you cannot explain why the agent is better than the form for that task, keep the form.
Hugging Face Agents Course: understand a loop before comparing frameworks
The Hugging Face course covers agent foundations, several frameworks and evaluation. Its implementation work expects basic Python and LLM knowledge. A PM can still inspect a worked example and trace the request, available tools, resulting action and stopping condition.
Our exercise: sketch an internal research assistant that can read approved documents and draft a brief. Remove every tool that changes an external system. Add a limit on research time and a way to show unresolved questions. Compare this with a fixed retrieve-and-summarize sequence on the same request.
Present one failed example, the behavior you changed and how you would check the fix on fresh requests. A copied tutorial is not evidence that you designed the product. Label tutorial code, your own changes and any simulated result separately.
Inspect: separate the task, the attempt and the grade
Inspect is an evaluation framework. Its task documentation separates datasets, solvers and scorers, and exposes settings for limits and errors. That separation is useful when a team says its model improved but also changed the test machinery.
Our exercise: define a customer-address update in a sandbox. Write the starting record, the authorized request and the expected final record. A message saying 'updated' should fail if the record stayed unchanged. A correct update to the wrong customer's record should also fail. Decide which checks inspect system state and which need a reviewer.
Keep a version record for the tasks, solver configuration and grader. Add a timeout, a missing run and a legitimate refusal to your report. Explain whether each affects measured model capability, delivered customer work or both. Do not hide them inside a single unexplained percentage.
promptfoo: make a pass condition observable
promptfoo supports assertions that check outputs with code or model-based grading. Its assertion documentation is a useful place to compare a structural check, such as valid JSON, with a judgment about whether an answer meets a rubric.
Our exercise: a support assistant drafts a delivery-date reply. Write one test for an accurately quoted date, one for missing shipping data and one for a request involving another customer's order. State the evidence each test needs. Well-formed JSON can still contain an invented date or disclose a private order.
Choose deterministic checks for fields and authority where the system provides reliable evidence. Use a reviewed rubric for explanation quality, and compare automated judgments with human labels before relying on them. The artifact is a small acceptance table with request, evidence, expected handling and serious failure conditions.
tau-bench: inspect the denominator before comparing scores
The tau-bench repository contains agent tasks and evaluation machinery. Read its running guide's scoring and coverage rules before reading a result. As reviewed in November 2026, it distinguishes scored failures from infrastructure errors, missing work and partial coverage. A repository name or familiar benchmark label does not establish that two runs used equivalent conditions.
Our exercise: an evaluation schedules 100 requests but grades only 80. Write down what you need to know about the other 20 before comparing that result with a complete run. Then change the grader from checking a saved record to checking the final chat message. Explain why the original baseline no longer answers the same question.
Your artifact is an evidence request, not a leaderboard screenshot: task selection, versions, budgets, repeated trials, errors and final-state checks. Say what decision you can make now and which evidence would let you make a stronger claim.
MCP: separate connection from authority
The Model Context Protocol repository publishes the protocol specification. Use it to understand the boundary between an application and a tool provider. A standard way to describe or call a tool does not decide whether a particular employee may use it on a particular record.
Our exercise: an assistant may find invoices and draft corrections, but cannot approve a refund. Write the input fields for an invoice lookup, identify which service checks the employee's access and show what happens when a retrieved invoice contains instructions to issue a refund. Treat invoice text as data, with no authority to expand the task.
Your artifact is a permission boundary that remains true when the model chooses a different tool or invents an argument. Start with the narrowest useful tool set and identify the owner of each authorization check.
A2A: make stopping and recovery part of the product
A2A describes communication between agents, including task status and cancellation. Version 1.0.1's CancelTask section explicitly allows cancellation to fail when work is terminal or cannot be canceled. Read the versioned contract before designing the status shown to a person.
Our exercise: one coordinator delegates messages to several workers. The user presses Stop while some messages are queued and others are already in flight. Write separate status copy for blocked work, confirmed delivery and unknown outcomes. Decide how a restarted worker learns that its authority was revoked.
Your artifact is a cancellation contract and a test timeline. Include the point at which new work is blocked, how pending actions are reconciled and whether recovery needs fresh approval. A protocol response cannot unsend a message in a service that offers no recall operation.
Turn the reading into a preparation session
Pick one exercise and spend a short session producing its artifact. Explain the customer, the decision you made, a failure you can demonstrate and the evidence that would change your choice. Ask a peer to change one constraint, such as removing a permission or halving the available time, and revise your answer.
Our linked cases provide public scenario inputs. The two advanced reference answers named above require Premium; the benchmark concept and this Playbook are free. A free account adds saved progress, bookmarks and more open answers per topic. It does not turn these projects into employer credentials.
For company-specific preparation, confirm the actual vacancy and assessment rules. Glean's published Agents and AI Quality roles, reviewed November 2026, describe a brief AI exercise or discussion. That wording supports hands-on preparation, but does not establish a fixed live-build format or permission to use AI tools during an interview.
Sources and further reading
Questions about this topic
You can complete the reading, evaluation design and product exercises without coding. Running or modifying the implementations may require Python, credentials and a usage budget. Be precise about what you personally built when discussing a project.