AI Engineer (LLM & Generative AI)
AI agents and tool calling questions
152 questions on AI agents and tool calling for ai engineer (llm & generative ai)s, easiest first.
Easy 45
- What the context window limitseasy
- What a tool definition containseasy
- The basic agent loopeasy
- What MCP is foreasy
- Workflow or agenteasy
- Reading the stop reasoneasy
- Describing tool parameterseasy
- What ReAct meanseasy
- MCP message formateasy
- What an MCP server offerseasy
- Choosing an MCP transporteasy
- Short-term and long-term memoryeasy
- Do tool definitions cost tokenseasy
- Limiting a parameter to fixed valueseasy
- What an idempotent tool iseasy
- Why trace an agent runeasy
- What a subagent iseasy
- The agent gets the date wrongeasy
- What lets an LLM take actionseasy
- Host, client and server in MCPeasy
- How a client finds MCP toolseasy
- Logging in to a remote MCP servereasy
- MCP compared with A2Aeasy
- Do we even need an agenteasy
- Handing over to a humaneasy
- Asking instead of guessingeasy
- More context, worse answerseasy
- Flagging a failed tool resulteasy
- Making a parameter mandatoryeasy
- Forcing the model to use a tooleasy
- Least privilege for agent toolseasy
- First step in improving an agenteasy
- A limit on loop turnseasy
- A tool that hangseasy
- The first MCP messageeasy
- What an MCP resource iseasy
- What a good tool description sayseasy
- Naming a tool parametereasy
- Processing invoices every nighteasy
- Planning a long task firsteasy
- Feedback for a coding agenteasy
- Users waiting on a slow agenteasy
- Testing tools on their owneasy
- A caller asks for an order addresseasy
- What agent benchmarks testeasy
Medium 60
- Who runs the toolmedium
- Two tool calls in one turnmedium
- Streaming tool-call argumentsmedium
- Debugging a bad agent answermedium
- A gate between chained stepsmedium
- Routing support ticketsmedium
- Voting with parallel callsmedium
- When subtasks are not known aheadmedium
- Evaluator and optimizer loopmedium
- A tool that returns 60,000 tokensmedium
- Writing a helpful tool errormedium
- Two tools both called searchmedium
- IDs the model keeps getting wrongmedium
- API error about a missing tool resultmedium
- GST totals from the modelmedium
- Meeting time from IST to Dubaimedium
- A report that takes 20 minutesmedium
- Where API keys should gomedium
- Loading files only when neededmedium
- What context compaction doesmedium
- Old tool results filling contextmedium
- Keeping progress across many hoursmedium
- Grading an agent fairlymedium
- User writes Dubai in Arabicmedium
- An MCP server asking for a completionmedium
- Server needs a detail from the usermedium
- Tools that appear after loginmedium
- Plain tools or an MCP servermedium
- Handoff or agent as a toolmedium
- Subagents doing duplicate workmedium
- Two subagents edit one filemedium
- Retrieval inside an agent loopmedium
- Thinking between tool callsmedium
- A clear way to finishmedium
- Who retries a flaky APImedium
- A payment charged twicemedium
- The model calls a missing toolmedium
- Prices in the wrong currencymedium
- Letting the agent pick output detailmedium
- Batch discounts for a chat agentmedium
- Four lookups, one at a timemedium
- A system prompt full of if-else rulesmedium
- Choosing examples for an agentmedium
- An optional parameter left outmedium
- Three tools for one bookingmedium
- Listing 40,000 contactsmedium
- A tool with tricky input formatsmedium
- Parallel calls that depend on each othermedium
- Agent crashes at step 37medium
- Previewing a bulk updatemedium
- Marking fetched content as datamedium
- Limiting where an agent can browsemedium
- One user running up the billmedium
- The current time in the system promptmedium
- What to track for a live agentmedium
- Asking the agent to review itselfmedium
- An empty list that was really an errormedium
- Telling a server which folders to usemedium
- Says it will call a tool, but does notmedium
- Whose permissions the agent usesmedium
Hard 47
- Injected instructions in a web pagehard
- Stuffing 100 chunks into the prompthard
- Agent stuck in a loophard
- Counting letters in a wordhard
- When a tool call failshard
- A suspicious third-party MCP serverhard
- Ten steps at 95 percent eachhard
- Reliability across three runshard
- Tool call cut off mid-argumentshard
- Where to enforce a refund caphard
- Three risky abilities togetherhard
- Cost of one agent taskhard
- An MCP server that breaks on logshard
- Trusting a readOnly hinthard
- A tool that changed after approvalhard
- Calling tools from codehard
- Is multi-agent worth ithard
- Tests passed, says the agenthard
- Agent changes the tests insteadhard
- Retries at three layershard
- Cheaper per call or per successhard
- Passing the client token onwardhard
- Users approving everythinghard
- An agent with a shell toolhard
- Noise between repeated eval runshard
- Testing an agent without live APIshard
- Cache misses in an agent loophard
- Debugging through a frameworkhard
- Input tokens across a long loophard
- Savings from caching a long prefixhard
- A safety check without extra delayhard
- Personal data in agent memoryhard
- Old memory, new addresshard
- Pausing to check policy mid-taskhard
- Screen clicking or an APIhard
- Cutting cost on easy agent stepshard
- Starting tools before the model askshard
- A subagent summary drops the key facthard
- A support ticket that gives ordershard
- A rule forgotten after compactionhard
- Is the agent cheaper than staffhard
- Blocking a bad agent changehard
- An agent asking for full Drive accesshard
- The same email sent three timeshard
- What an approval prompt should showhard
- Traces full of customer datahard
- Agent drifts from the goalhard