Oliver Cheng (Oli)
← Back to blog
Oli Cheng 3 min read AI Engineering

From Vibe Coding to Agent Org Design

We went from copy-paste coding in 2023 to managing persistent AI workers in 2026. The next skill is organizational design for agent teams.

  • AI Engineering
  • Agents
  • Organizational Design
  • Developer Tools
  • Product Strategy
From Vibe Coding to Agent Org Design

The coding loop changed fast.

In early 2023 and much of 2024, “vibe coding” mostly meant pasting code from ChatGPT into an editor, crossing fingers, and patching errors manually until things stopped breaking.

That phase worked, but it was noisy:

  • weak state management,
  • lots of context loss,
  • brittle integration,
  • no stable execution memory.

You were still doing almost all the orchestration yourself.

Quick definitions (plain English)

  • Vibe coding: trying to build mainly by prompting, with limited architecture and test discipline.
  • Orchestration: deciding task order, context handoff, and ownership across steps.
  • Persistent AI worker: an assistant process that can continue a task across many steps rather than one reply.
  • Agent org design: structuring multiple AI workers into clear roles and review loops.

Phase 1: copy-paste optimism

The first wave was about possibility.

You could get real acceleration if you already knew how to debug. But if you did not understand architecture, type constraints, and dependency surfaces, you were basically assembling random parts and hoping the machine turned on.

Most of the gain was in draft speed, not delivery reliability.

Phase 2: IDE copilots and tighter loops

By late 2024, tools like Cursor and IDE-native assistants improved the loop significantly.

You got:

  • better in-file context,
  • project-aware suggestions,
  • faster refactors,
  • less copy/paste churn.

This was a material jump. But it was still a one-operator workflow: one human, one tool, one context window, one chain of edits.

Phase 3: persistent AI workers

Now with Codex/Claude-style agent workflows, the model is no longer just autocomplete. It is closer to a worker that can continue operating as long as it has clear constraints and useful tasks.

That flips the bottleneck.

The limiting factor is no longer how fast you can type. It is how well you can:

  1. define outcomes,
  2. break work into coherent chunks,
  3. provide the right context,
  4. review and steer decisions.

In other words, coding is becoming management.

The next step: managing teams of agents

The natural next iteration is multi-agent organizational design.

Not one assistant, but a system of specialized roles:

  • planner,
  • implementer,
  • reviewer,
  • tester,
  • documentarian,
  • monitor.

At that point, your core skill is to communicate:

  • priorities,
  • goals,
  • context boundaries,
  • escalation rules,
  • ownership lines between agents.

That looks a lot like running a team.

Why MSCI 211 suddenly matters

When I took organizational design (MSCI 211) at Waterloo for my cognitive science minor, I expected it to help with human teams.

It still does. But the surprising 2026 twist is that the same principles now apply to agent teams:

  • role clarity prevents overlap,
  • communication channels prevent stale assumptions,
  • incentive design (or objective design) drives behavior,
  • feedback loops determine system quality over time.

The org chart is becoming executable.

The talent tree argument

Tech careers are a skill tree, not a linear ladder.

You never see the full tree at once. You only see the next two or three unlocks.

In 2023, one unlock was “prompting.” In 2024, another unlock was “tooling fluency.” In 2025-2026, the unlock is “agent orchestration and system supervision.”

The people who adapt fastest are usually not those with the loudest takes. They are the ones who keep redistributing effort to the next constraint.

Right now, that constraint is organizational intelligence.

Practical operating model

If you are building with agents today, run a lightweight structure:

  1. Mission layer: define user outcome and non-negotiables.
  2. Role layer: assign explicit responsibilities per agent.
  3. Context layer: bound what each role can see and change.
  4. Review layer: require periodic synthesis and conflict checks.
  5. Ship layer: one final integrating owner (human).

The human remains accountable for product truth and ethical direction.

Closing

The future of software work is not “humans disappear.” It is “humans who can design and steer systems become much more leveraged.”

The interface changed from keyboard-only production to manager-level orchestration.

That is not the end of engineering. It is engineering at a higher organizational layer.

Related notes:

Still learning the next unlock with you, Oli
March 3, 2026