Maddy Zhang uses one budgeting app to explain three layers of agent engineering in this ten-minute tutorial.

Make one agent’s work checkable

  • A harness is the software around a model: its tools, context, project rules and feedback.
  • For a bug that double-counts recurring expenses, give the agent a test command and block completion until the required checks pass.
  • Turn repeated failures into enforceable checks. Passing tests still depends on whether those tests cover the real requirement.

Give repeated work a trigger and a stopping point

  • A loop finds work, starts an agent and repeats until a checkable condition is met.
  • Zhang’s maintenance example combines scheduled discovery, isolated Git worktrees, reusable instructions, tool connections and a separate reviewing agent.
  • Save progress outside the conversation so the next run can distinguish finished work from pending work.
  • Define “done” as specific tests and checks passing. Keep human review of proposed changes to limit comprehension debt: code accumulating faster than people understand it.

Coordinate parallel work only where it fits

  • A graph connects work steps with rules for dependencies, parallel execution, review and retries.
  • For a JavaScript-to-TypeScript migration, map the files, divide the work, review each result and run the combined test suite.
  • Graph nodes can contain ordinary code, one model call or an entire agent. Loops and graphs overlap as control structures.
  • LangChain’s research example favors an adaptive agent loop when the next steps cannot be predicted in advance.

Pay for the extra machinery deliberately

  • Start with a harness for one bug, a loop for recurring maintenance and a graph for work that divides into independent pieces.
  • Zhang cites Anthropic’s research-system data: agents used about 4× the tokens of chat interactions, and multi-agent systems about 15×.
  • Those are workload-specific token observations, not universal multipliers for runtime or dollar cost.
  • Disclosure: Granola sponsors the meeting-context/MCP segment; its connector illustrates bringing outside information into an agent’s tools.

“The skills that carry across all three layers are defining done, breaking work down and verifying the results.” — Maddy Zhang, 9:30–9:36