Question 1

How do you detect hallucination in code the agent wrote?

Accepted Answer

Answer in layers, cheapest first: 1. Compile — catches non-existent APIs immediately. The cheapest gate. 2. Lint — flags deprecated or incorrect usage. 3. Tests — catch code that compiles but behaves wrongly. 4. Review — catches logic and architecture errors. 5. Official docs — for concrete factual claims (API level, default value, behaviour). Then add the key nuance: none of these layers is AI-specific — they all exist anyway. What changes is that with more code volume the gates must be automated; manual review no longer scales.

Question 2

You need to work with a newly released Jetpack library the model does not know. What do you do?

Accepted Answer

Start from the principle: anything the model does not know must be brought into context. Concrete steps: 1. Have the agent read the relevant official docs page (if it has web access), or paste the essential part yourself. 2. Bring the library's official sample into the project and point at it as the pattern. 3. Start with a small spike module — let the agent try the API there and compile before touching the main code. 4. Record the learned pattern in CLAUDE.md or a skill so you don't re-explain it next time. Otherwise: the model extrapolates from similar older APIs and writes plausible but wrong code that stays hidden until compile time.

Question 3

Which Android tasks should not be delegated to an agent, in your view?

Accepted Answer

Rather than a hard "never" list, propose criteria: the cost of a mistake and verifiability. Areas I don't fully delegate: Cryptography and key storage — a mistake compiles, passes tests, and the weakness stays. Payment and financial logic — regulatory requirements, very high cost of error. Collection/sharing of user data — privacy and legal requirements. ProGuard/R8 and release configuration — mistakes only show up in production. But even these can be done with the agent: it proposes options, a human decides and documents. The difference is that the agent is an assistant, not the author; final responsibility stays with the human.

Risk area	Typical Android symptom	How to catch it
Non-existent API	A Compose modifier or parameter that does not exist	Compilation — it fails immediately
Outdated approach	A deprecated library or old Gradle syntax	Lint warnings plus a check against the official docs
Wrong behaviour (it compiles)	A coroutine launched in the wrong scope	Unit/UI tests, code review
Fabricated fact	"This exists from API level 26" — an unverified figure	Checking developer.android.com

Limitations: hallucination and knowledge cutoff

📚 Sources and documentation