Eric names three skills that separate the top 1% in an agentic workflow 8:23:

  1. Thinking about problems in piecemeal fashion
  2. Knowing how to context switch among parts of the problem and among agents
  3. Knowing how to develop agent-friendly codebases

Skill 1 — Piecemeal problem decomposition

  • "The best engineers are the ones that know how to take a very complex task and then break it down to the subtasks that need to get organized" 8:45.
  • Decomposition includes translating the dependencies between subtasks, not just listing them 8:45.
  • Only from that place can you kick work off to agents 8:45.

Skill 2 — Context switching across agents

The concrete loop Eric describes 9:05:

flowchart TD
    D[Decompose the task] --> P1[Assign chunk 1 + spend time making its plan]
    P1 --> W1[Agent 1 works]
    W1 --> T{"Right moment to switch?"}
    T -->|yes| P2[Kick off chunk 2 to agent 2]
    P2 --> T2{"Right moment again?"}
    T2 -->|yes| P3[Add a task to agent 3's bucket]
    P3 --> S["Supervise: who is stuck?<br/>who needs clarification?"]
    S --> T
  • The hard part is knowing the right point to switch to another agent and kick off the next piece 9:05.
  • Moving seamlessly between multiple working agents is "a pretty complex skill because it's a bit like a psychological skill" — tracking what's going on with each agent, where it's stuck, where it needs more clarification 9:25 9:42.
  • The mental model both parties land on: "You're the scrum master of a team of agents" — the manager who has to maintain all the context, because agents are "the most eager interns on the planet." They do exactly what you ask, but they need supervision 9:42 10:00.

Skill 3 — Agent-friendly codebases

Eric flags this as "not talked about enough" 10:00. The principle: set your agents up for success.

Concrete requirements named:

  • Strong test coverage in the codebase 10:00
  • A strong build system 10:00

The analogy: the house is built before someone enters it and starts adding a room on top 10:20.

  • Payoff: only once that foundation exists can you confidently run multiple agents in parallel without them producing sloppy code, trusting they'll build in the context of the foundation you set up 10:20.