Skip to content

Instantly share code, notes, and snippets.

@soasme
Created May 12, 2026 18:46
Show Gist options
  • Select an option

  • Save soasme/91dfb8f49c48678e5560a637676eadc5 to your computer and use it in GitHub Desktop.

Select an option

Save soasme/91dfb8f49c48678e5560a637676eadc5 to your computer and use it in GitHub Desktop.

The New Engineering Teams

How AI coding agents change the way teams should work


Opening

Most engineering orgs were designed for one constraint:

Code was expensive to produce.

That constraint shaped everything:

  • planning
  • review
  • ownership
  • hiring
  • documentation
  • management layers

Now AI coding agents are changing the cost structure.

So the question is no longer:

“How do we write code faster?”

It is:

“What breaks when code becomes abundant?”


The Big Idea

AI does not remove bottlenecks.

It relocates them.

The new bottlenecks are less about typing code and more about:

  • deciding what should exist
  • checking whether it works
  • judging whether it is safe
  • aligning people around tradeoffs
  • maintaining quality over time

A Useful Mental Model

Old engineering org:

Protect scarce engineering time.

AI-native engineering org:

Protect scarce human judgment.

That one shift changes nearly every team ritual.


The Old World

When implementation was slow, it made sense to invest heavily before writing code:

  • long roadmaps
  • architecture debates
  • design docs
  • estimation meetings
  • strict component ownership
  • multi-stage reviews
  • careful prioritization

The system was optimized to avoid wasting developer effort.


The New World

When implementation becomes cheaper, the cost center moves.

Now the expensive parts are:

Ambiguity

Verification

Coordination

Risk

Taste

Maintenance

A team can create more code than it can responsibly understand, test, and operate.


The Hidden Risk

AI-assisted teams may feel faster immediately.

But speed can create invisible debt:

  • too many PRs
  • too much generated surface area
  • shallow understanding
  • unclear ownership
  • weak tests
  • more regressions
  • noisier reviews

The goal is not more code.

The goal is safer learning velocity.


Reframe the Engineering Manager’s Job

The manager’s job shifts from:

“How do I allocate limited coding capacity?”

to:

“How do I design a system where fast output stays high quality?”

That means managing constraints, feedback loops, and judgment points.


What Should Change First?

Do not start by adding an AI policy.

Start by auditing team rituals.

For each workflow, ask:

Was this created because coding used to be expensive?

If yes, it may need to be removed, automated, or redesigned.


The Three-Way Process Test

Every recurring workflow should face one of three outcomes:

Delete it

If it no longer protects anything important.

Automate it

If it is repetitive, mechanical, or summary-based.

Keep it lighter

If human judgment is still required.


Planning Becomes More Experimental

Old planning assumption:

We need to think carefully because building is costly.

New planning assumption:

We can learn faster by building small versions.

This does not mean “no planning.”

It means planning becomes more just-in-time.


Roadmaps Should Become Less Fragile

A long roadmap can still provide direction.

But it should not pretend the future is fixed.

In an AI-native team, discovery accelerates:

  • prototypes appear faster
  • internal tools evolve faster
  • customer feedback can be acted on faster
  • ideas can be tested before they are fully debated

So planning should become more adaptive.


Replace Some Debates with Artifacts

Instead of spending a week arguing about an approach:

Generate two or three implementation options.

Then compare them concretely:

  • Which API is cleaner?
  • Which migration is safer?
  • Which version is easier to test?
  • Which one creates less long-term complexity?
  • Which one fails more gracefully?

Real code can make abstract debates shorter.


But Do Not Let Code Win by Default

There is a danger:

The easiest generated implementation becomes the accepted implementation.

That is not good engineering.

AI makes options cheap.

It does not make decisions automatic.

Human judgment still chooses.


The New Planning Principle

Prototype quickly.

Decide deliberately.

Verify aggressively.

That is the AI-native planning loop.


Verification Moves to the Center

When output increases, review alone cannot scale.

The team needs stronger automatic checks:

  • tests
  • linting
  • static analysis
  • type checks
  • security scanning
  • regression suites
  • preview environments
  • eval harnesses
  • CI quality gates

AI-generated code should meet a higher bar, not a lower one.


“Shift Left” Becomes Mandatory

If AI helps produce more code earlier, quality checks must also happen earlier.

Do not wait for a senior engineer to discover obvious issues in review.

Let automation catch:

  • formatting
  • simple bugs
  • missing tests
  • broken builds
  • risky dependencies
  • repeated patterns
  • obvious regressions

Humans should review what machines cannot judge well.


Code Review Gets Repositioned

Traditional code review often mixes many jobs:

  • style enforcement
  • correctness checks
  • knowledge sharing
  • architecture review
  • risk review
  • product review
  • mentoring

AI and automation can absorb part of this.

Human review should become more focused.


What AI Can Handle in Review

AI can help with:

  • summarizing PRs
  • finding duplicated logic
  • checking style
  • proposing tests
  • responding to simple comments
  • spotting obvious edge cases
  • explaining unfamiliar code
  • updating docs or comments

This reduces review noise.


What Humans Must Still Own

Humans should stay responsible for:

Product judgment

Security boundaries

Legal and compliance risk

Architectural direction

Customer impact

Long-term maintainability

Taste

These are not just code questions.

They are responsibility questions.


Review Philosophy

Do not ask humans to inspect every small mechanical detail.

Ask humans to answer:

Should this change exist?

Is this the right shape?

What could go wrong?

Are we comfortable owning it?

That is a much better use of scarce attention.


Ownership Needs a New Vocabulary

In AI-assisted development, “who wrote this?” becomes a weak question.

A better ownership model asks:

  • who understands this area?
  • who approved the direction?
  • who can debug it?
  • who can explain it to customers?
  • who is accountable for the outcome?
  • what context did the agent use?

Authorship matters less.

Accountability matters more.


Code Ownership Becomes Context Ownership

The important owner may not be the person who typed the code.

The important owner is the person who can say:

“I understand why this exists, how it works, and what risk it introduces.”

That is the ownership standard.


Cross-Functional Work Speeds Up

AI also changes boundaries between roles.

Engineers can draft copy.

Designers can propose small UI patches.

PMs can inspect implementation options.

Support feedback can be summarized automatically.

Research can be turned into tickets faster.

The handoffs shrink.


Specialists Still Matter

AI does not remove designers, PMs, security experts, or content specialists.

It reduces waiting.

A good pattern is:

AI drafts.

Specialist judges.

Team ships faster.

The human expert becomes more leveraged.


Example Pattern

Before:

Bug fixed
→ wait for copy
→ wait for review
→ update ticket
→ ship later

After:

Bug fixed
→ AI drafts copy
→ human approves or edits
→ tests pass
→ ship

The improvement is not “AI replaced content design.”

The improvement is “the team removed idle time.”


Hiring Criteria Changes

If AI raises baseline coding output, hiring should not over-index on raw implementation speed.

Look harder for:

Product-minded builders

People who can find useful problems, shape solutions, and care about experience.

Deep technical experts

People who can reason about complex systems, reliability, security, performance, and failure modes.


What Becomes Less Differentiating

These become less special by themselves:

  • typing lots of code
  • implementing straightforward CRUD
  • translating tickets into code
  • fixing obvious bugs
  • producing boilerplate quickly

They still matter.

They are just less rare.


What Becomes More Valuable

These become more valuable:

  • asking better questions
  • decomposing ambiguous work
  • spotting risk early
  • designing verification
  • making taste calls
  • simplifying systems
  • aligning stakeholders
  • operating production systems
  • knowing when not to build

Team Structure Should Flatten

If every decision requires climbing a hierarchy, AI-generated speed gets trapped.

Flatter teams help because:

  • context moves faster
  • people can self-serve more
  • managers stay close to the work
  • ownership is clearer
  • experiments require less ceremony

The org should not add management layers just because output increased.


Managers Need Hands-On Context

In an AI-native team, a manager who only manages process may lose touch quickly.

The healthiest pattern:

Managers retain some IC-style exposure.

They should use the tools, inspect the work, understand the workflows, and feel the friction themselves.


Management Becomes System Design

Managing an AI-native engineering team is closer to designing a production system.

You need:

  • fast feedback loops
  • clear ownership
  • safety checks
  • observability
  • failure handling
  • escalation paths
  • continuous improvement

The team itself becomes the system.


Documentation Changes Role

Documentation used to be the main way people learned a system.

Now code, tests, PRs, tickets, and agent-readable context can become the living knowledge base.

The principle:

Keep knowledge close to the source of truth.


Code as the Knowledge Base

If the code is the most accurate artifact, make it easier for agents and humans to inspect.

Useful practices:

  • keep specs in the repo
  • write clear tests
  • maintain architecture notes near implementation
  • use machine-readable conventions
  • let agents check whether docs drift from code

Docs still matter.

Stale docs matter less than verified context.


Onboarding Gets Rewritten

New team members can ask AI unlimited “basic” questions:

  • What does this service do?
  • Where is the auth logic?
  • Why is this test failing?
  • What changed in this PR?
  • How do I run this locally?
  • What are the risky parts of this module?

This lowers the social cost of learning.


Knowledge Sharing Becomes Pull-Based

Old knowledge sharing:

Senior people push context through meetings and docs.

New knowledge sharing:

People and agents pull context when needed.

This works only if the repo, tickets, and history are structured well enough to be queried.


Status Reporting Should Be Questioned

Many status rituals exist because managers could not easily see work in progress.

But if the signals already exist in:

  • commits
  • PRs
  • issues
  • docs
  • support tickets
  • incident reports
  • chat threads

then AI can summarize them.

Do not make humans manually restate what the system already knows.


Metrics Need Care

Useful signals:

  • onboarding ramp time
  • PR cycle time
  • review latency
  • escaped regressions
  • incident rate
  • test coverage quality
  • deployment frequency
  • customer-impacting defects
  • AI-assisted workflow adoption

But measure outcomes, not theater.


The Vanity Metric Trap

A bad metric:

“What percentage of our code was written by AI?”

That number may go up while quality goes down.

Better questions:

  • Are we shipping better?
  • Are regressions decreasing?
  • Are reviews faster?
  • Are engineers less blocked?
  • Are customers happier?
  • Are systems easier to operate?

The Real Goal

The goal is not to maximize AI usage.

The goal is to maximize:

learning speed

product quality

engineering reliability

team leverage

AI usage is only useful if it improves those.


The New Flywheel

AI reduces implementation cost.

Lower implementation cost enables more experiments.

More experiments require stronger verification.

Stronger verification enables safer shipping.

Safer shipping creates faster learning.

Faster learning makes the team more competitive.


What Can Go Wrong

AI-native engineering can fail if teams:

  • generate too much unreviewed code
  • weaken ownership
  • treat tests as optional
  • let agents create inconsistent patterns
  • hide behind automation
  • skip product judgment
  • reward volume over quality
  • confuse speed with progress

The danger is not that AI writes code.

The danger is unmanaged abundance.


Open Questions for AI-Native Orgs

Some questions do not yet have clean answers:

  • Should teams still be organized by platform?
  • How much review can be automated safely?
  • How do people feel ownership when agents contribute everywhere?
  • How should performance be evaluated?
  • What does seniority mean when implementation is cheap?
  • How do we prevent generated complexity?

These are management problems, not just tooling problems.


The Practical Starting Point

Pick one painful workflow.

Examples:

  • weekly status updates
  • bug triage
  • design doc review
  • PR review
  • onboarding
  • release notes
  • support summary
  • incident follow-up
  • test generation

Do not transform everything at once.

Start with the noisiest workflow.


Workflow Audit Questions

Ask:

Why does this workflow exist?

Then:

  • What bottleneck was it designed for?
  • Is that bottleneck still real?
  • What part is mechanical?
  • What part requires judgment?
  • Could AI draft it?
  • Could automation verify it?
  • Could we delete it entirely?

A Better Rule of Thumb

Use AI for:

generation

summarization

comparison

cleanup

first drafts

mechanical checks

context retrieval

Use humans for:

direction

taste

accountability

risk

prioritization

final judgment


Final Message

AI-native engineering is not just adopting a coding assistant.

It is redesigning the team around a new scarcity.

Code is less scarce.

Judgment is more scarce.

The winning teams will not be the ones that generate the most code.

They will be the ones that turn abundant code into reliable product learning.


Closing Slide

Do not ask:

“How much faster can we code?”

Ask:

“What should humans stop doing, so they can focus on what only humans should decide?”

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment