Skip to content

Interview index: 232 source-backed questions

An index of 220 review questions from 22 architecture chapters, plus 12 checks from the two engineering-method chapters. Every entry leads back to a worked answer and source trail.

Four ways to use this page:

  1. By chapter order: read chapters 1 through 22. Each has 10 questions, and later chapters go deeper.
  2. By topic cross-cut: use the topic quick-find table to see the same question class from different angles.
  3. As an interview question bank: every chapter link lands on its appendix, where the complete questions, answers, and follow-ups live.
  4. As a self-checklist: when building your own agent, find the chapter matching the design choice ahead, then audit the decision against those 10 questions.

Themes: harness definition, agent loop, context, tool, verifier. These 50 cover the vocabulary and trade-offs of agent fundamentals.

ChapterTopicLink
§1Agent Harness definition, concept boundaries, picking among the four systems01 · Overview
§2Agent loop five phases (perceive / plan / act / observe / verify), turn data structure, prompt assembly, context compaction, error recovery, tool dispatch, verifier02 · Agent Loop
§3Context systems: full spectrum from markdown to 10-layer assembly; how prefix cache holds up03 · Context system
§4Tool system: protocol, registration, parallelism, security gating04 · Tool system
§5Verifier: hard / soft / give-up tiers + stop detection05 · Verifier

Themes: real coding-agent engineering details. Patch algorithm, shell invocation, git workflow, code review automation. These tell you what extra trouble a coding agent has over a chat agent.

ChapterTopicLink
§6File edit + patch algorithm: apply_patch V4A multi-version diff merge06 · File edit patch
§7Shell execution: execpolicy / sandbox / approval mode07 · Shell execution
§8Git workflow: worktree / branch / commit automation08 · Git workflow
§9Code review: /security-review + LLM-as-reviewer pattern09 · Code review

Themes: subagent, session lifecycle, permissions, sandbox, multi-channel entry. These 50 are the engineering foundation for keeping the agent stable in production.

ChapterTopicLink
§10Subagent: fork / inline boundary, context isolation10 · Subagent
§11Session lifecycle: start / resume / interrupt / persist11 · Session lifecycle
§12Permissions and approvals: AskForApproval 4 levels + autoMode rules + INSTALL_POLICY 12-cell matrix12 · Permissions
§13Sandbox execution: macOS seatbelt / Linux bwrap+seccomp+landlock / Windows13 · Sandbox execution
§14Multi-channel entry: CLI / TUI / IDE / web / cron14 · Multi-channel entry

Evolution, safety, and long-running systems

Section titled “Evolution, safety, and long-running systems”

Themes: observability, memory, skills, scheduled tasks, self-improvement, security, todo lists, execution-state surfaces. These 80 are the core for long-running agents that stay recoverable and safe.

ChapterTopicLink
§15Observability and cost: token / latency / tool.loop detection15 · Observability and cost
§16Memory system: 4 MemoryTypes, frozen snapshot, decay, prefix cache16 · Memory
§17Skill system: 8 crates / skillify / skill-scanner / 12-cell trust matrix17 · Skills
§18Cron and background tasks: tick + scheduler lock + isolated-agent18 · Cron background tasks
§19Self-improvement: Phase1+2 / skillify / passive indexing / in-turn explicit writes19 · Self-improvement
§20Security: sandbox / external-content / redact / cosign provenance20 · Security
§21Todo List: update_plan / TodoWrite / runtime progress / compaction recovery21 · Todo List
§22Execution State Surfaces: Plan / Todo / tool progress / task progress / terminal status / away summary routing22 · Execution State Surfaces

Topics: Loop Engineering and Graph Engineering. These are portable methods rather than four-system comparisons: how budgets are set, what edges carry, and where guardrails live. The material is grounded in mini-swe-agent, smolagents, crewAI, and open-source Codex.

ChapterTopicsLink
§23Loop engineering: five components, 3-dimension budgets, grace call, error taxonomy, breaker, trajectory, verifier hardness23 · Loop Engineering
§24Graph engineering: four node types, five edge elements, permissions-only-narrow, split decision, freezing, runaway protection24 · Graph Engineering

This section re-slices the 220 questions from the architecture chapters; the method chapters’ 12 checks remain in the section above. The same engineering problem appears from different angles in different chapters, so horizontal reading is often faster.

Aligning vocabulary across the four systems. Common in “what’s the difference between A and B” or “why this split”.

SourceQuestion
§1 Q1Agent Harness vs Agent model
§2 Q1Is “turn” a physical or logical unit
§4 Q1Tool / Function / Skill boundary
§10 Q1Subagent / fork / inline distinction
§17 Q1Skill / prompt template / tool / agent boundary
§18 Q1Cron job / background task / long-running task differences
§21 Q1Todo list vs Plan Mode boundary
§22 Q1Execution-state surface vs todo-list boundary

Why not B, with concrete context. Common in “why did Codex use X while Hermes used Y.”

SourceQuestion
§2 Q4 / §3 Q6When to use prefix cache vs recompute
§11 Q3Whether session persistence needs versioning
§13 Q5Why sandbox doesn’t ban all network
§15 Q4How observability data avoids exploding disk
§16 Q5Memory limit in tokens vs chars
§18 Q3Cron failure auto-disable vs retry
§19 Q3OpenClaw halfLifeDays=30 derivation
§20 Q3fail_open vs fail_closed
§21 Q6Lightweight TodoWrite vs file-backed Tasks
§22 Q6Why progress display needs per-platform settings

Concrete APIs, data structures, source pointers. Common in “why does this code do that” / “how does it handle X at runtime.”

SourceQuestion
§2 Q7 / §6 Q4apply_patch V4A multi-version diff merge
§3 Q8Hermes 10-layer prompt assembly ordering
§4 Q6dispatchToolUseBlocks parallel dispatch implementation
§5 Q5Codex goals.rs convergence judgment
§12 Q4AskForApproval 4 levels mapping to UI
§16 Q8Hermes _file_lock fcntl/msvcrt cross-platform
§17 Q6Hermes 12-cell INSTALL_POLICY calculation
§20 Q1OpenClaw random 8-byte ID wrapping
§21 Q8Todo replace vs merge API trade-off
§22 Q3MCP progress vs tool result protocol boundary

Injection / poisoning / secrets / supply chain. Common in “how does this attack get defended” / “why design it this way.”

SourceQuestion
§12 Q9Keeping user password out of prompt
§13 Q7Common sandbox-escape paths
§17 Q7OpenClaw skill-scanner 3 severities
§19 Q5Hermes reasoning for blocking invisible unicode
§19 Q106-layer security defense stack
§20 Q1external-content random ID forgery prevention
§20 Q2tirith exit code as source of truth
§20 Q5Codex memory prompt explicit “treat as data”
§20 Q105-layer general security defense
§21 Q9Why todo completion needs verification nudges
§22 Q8Why progress surfaces still need no-progress detectors

How to actually build your agent. Common in “what’s the minimum production needs” / “what’s the minimum viable implementation.”

SourceQuestion
§1 Q9Scenarios where heavy harness is wrong
§5 Q105 iron rules for verifier design
§10 Q10General subagent design checklist
§15 Q10General observability framework
§16 Q10General 5-layer memory architecture
§17 Q10General 6-layer skill system
§18 Q10General 7-layer cron system
§19 Q10General 6-layer self-improvement defense
§20 Q10General 5-layer security defense
§21 Q10Minimum 5-layer todo/progress surface
§22 Q10Minimum execution-state router fields

After 220 questions, the next step is turning insight into reusable engineering capability. Two paths:

  • Build directly: pick a chapter’s implementation recipe and follow it to a minimum viable version.
  • Crystallize into a skill: use the companion ~/.claude/skills/build-your-own-agent/ skill to invoke the best practices distilled from these 220 questions on demand. See the <BuildRecipe /> block at the end of each chapter.