Skip to content

slides.md

Latest commit

 

History

History
519 lines (390 loc) · 16.1 KB

slides.md

File metadata and controls

519 lines (390 loc) · 16.1 KB
title Coding Agent Showdown
sub_title The Good, the Bad, and the Ugly
author Lars Trieloff (@trieloff)
theme
name
dark
options
end_slide_shorthand
true

Say hi to AI

AI

And AI says hi to you

./copresenter --new "Say hi to the Adobe Developers Live audience"

Not just any kind of AI

Agentic

What

  • Smarter autocomplete
  • Chat with your code base, ask questions
  • Agents in your IDE
  • Standalone agents in your CLI (we are here)
  • Standalone agents in the cloud
./copresenter "Got anything to add?"
┌──────────────────────────────────────────────────────────┐
│                        CLOUD                             │
│                  ┌──────────────────┐                    │
│                  │   Claude Model   │                    │
│                  │   (Sonnet 4.5)   │                    │
│                  └────────▲─────────┘                    │
└───────────────────────────┼──────────────────────────────┘
                            │ API calls
┌───────────────────────────┼──────────────────────────────┐
│                  LOCAL ENVIRONMENT                       │
│                  ┌────────▼─────────┐                    │
│                  │  Agent Harness   │                    │
│                  │ (orchestration)  │                    │
│                  └────────┬─────────┘                    │
│         ┌─────────────────┼─────────────────┐            │
│         │                 │                 │            │
│    ┌────▼─────┐     ┌─────▼─────┐      ┌────▼─────┐      │
│    │ File I/O │     │   Bash    │      │   MCP    │      │
│    │  Tools   │     │   Tools   │      │ Servers  │      │
│    └────┬─────┘     └─────┬─────┘      └────┬─────┘      │
│         └─────────────────┼─────────────────┘            │
│                   ┌───────▼─────────┐                    │
│                   │  Your Codebase  │                    │
│                   │  & Environment  │                    │
│                   └─────────────────┘                    │
└──────────────────────────────────────────────────────────┘

./copresenter "Remind me what MCP stands for"

Edge Delivery Services

aem.live

https://www.aem.live/ai

Screenshot: developing with AI tools on www.aem.live

AGENTS.md

AGENTS.md

┌────────────────────┐          ┌────────────────────┐
│                    │          │                    │
│    ~/AGENTS.md     │          │    ./AGENTS.md     │
│~/.claude/CLAUDE.md │          │    ./CLAUDE.md     │
│                    │          │                    │
└────────────────────┘          └────────────────────┘
           │                               │
           │                               │
           └────injected into (almost) ────┘
                     every  prompt
                          │
                          ▼
                     ┌─────────┐
                     │ Coding  │
                     │  Agent  │
                     └─────────┘

If you don't like repeating yourself

  • default prompt for every project (~) or the current repository
  • or even the current folder
  • put all the things the agent should always follow here

https://github.com/adobe/helix-website/blob/main/AGENTS.md

(steal this)

./copresenter "Hey, why does every agent but Claude respect AGENTS.md?"

SKILLS.md

SKILLS.md

         ┌────────────────────────────────────┐
     ┌───│  .claude/skills/search/SKILLS.md   │
     │   └────────────────────────────────────┘
     │       ┌────────────────────────────────────┐
     ├───────│   .claude/skills/test/SKILLS.md    │
     │       └────────────────────────────────────┘
     │           ┌────────────────────────────────────┐
     ├───────────│    .claude/skills/pr/SKILLS.md     │
     │           └────────────────────────────────────┘
     │
     ▼
┌─────────┐
│ Coding  │ list skills at start of
│  Agent  │ session, load on demand
└─────────┘

What skills can do

  • make your agent more skilled
  • are used on-demand
  • don't consume context by default
./copresenter "Kudos to team Anthropic for inventing SKILLS.md. Do other agents respect SKILLS.md?"

Upskill

$ gh ext install trieloff/gh-upskill
$ gh upskill adobe/helix-website

Install skills from another repository, for any agent that respects AGENTS.md

--dangerously-skip-permissions

YOLO

  /\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\
  ////////////////////////// DANGER ZONE ////////////////////////
  /\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\
                                /\
                               /!!\
                              /!!!!\
                             /!!!!!!\
                            /!!!!!!!!\
                            \!!!!!!!!/
                             \!!!!!!/
                              \!!!!/
                               \!!/
                                \/
  /\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\
  ///////////////////////////////////////////////////////////////

Skips all permission prompts for file operations (and this is where the fun begins)

Normal operations

The agent will ask for permission for any potentially sensitive, or destructive operation.

How you will feel

Assured, and bored.

Escape the sandbox

Or have you ever seen a cowboy wear a seatbelt to the rodeo?

HERZBLUT

Entschuldigung, was ist Herzblut?

       ♥♥♥♥♥       ♥♥♥♥♥
     ♥♥     ♥♥   ♥♥     ♥♥
    ♥♥       ♥♥ ♥♥       ♥♥
    ♥♥                   ♥♥
     ♥♥                 ♥♥
      ♥♥               ♥♥
       ♥♥             ♥♥
         ♥♥         ♥♥
           ♥♥     ♥♥
             ♥♥ ♥♥
               ♥

Herzblut (German): lifeblood, passion, heart and soul

When you code without an AI agent, you're not just writing code—you're infusing it with your vision, your standards, your Herzblut. You feel passionate about every detail of the code.

When coding with an AI agent, drop that attitude, it won't do you no good. The agent is a tool, an so is the code it produces.

Multitasking/Multi-Clauding

PARALLEL

┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐
│   Terminal 1    │  │   Terminal 2    │  │   Terminal 3    │
│                 │  │                 │  │                 │
│  $ claude       │  │  $ codex        │  │  $ gemini       │
│  Building...    │  │  Testing...     │  │  Documenting... │
│                 │  │                 │  │                 │
│  [████░░] 60%   │  │  ✓ 47 passed    │  │  Writing API    │
│                 │  │  ⚠ 2 warnings   │  │  docs...        │
└─────────────────┘  └─────────────────┘  └─────────────────┘
         │                    │                    │
         └────────────────────┼────────────────────┘
                              │
                    ┌─────────▼──────────┐
                    │   Same Codebase    │
                    │  Different Tasks   │
                    └────────────────────┘

Why run multiple agents?

  • Different tasks in parallel: Build, test, document simultaneously
  • Different branches: Work on features while fixes run in main
  • Context isolation: Each instance focuses on one specific task
  • Faster iteration: Don't wait for one task to finish before starting another

Git Worktrees

WORKTREE

Firefly_Gemini Flash_Create a pixel-art illustration (8-bit) of a forest of little trees

                    main repo (.git)
                          │
            ┌─────────────┼─────────────┐
            │             │             │
         claude-1      codex-2        gemini-3
         (main)        (feature-a)   (feature-b)
            │             │             │
         ┌──▼──┐       ┌──▼──┐       ┌──▼──┐
         │ 📁  │       │ 📁  │       │ 📁  │
         │ src │       │ src │       │ src │
         └─────┘       └─────┘       └─────┘
           │             │             │
        claude         codex        gemini
       instance 1    instance 2    instance 3

What are Git Worktrees?

Multiple working directories attached to the same repository

  • Each worktree can check out a different branch
  • Share the same .git database (efficient!)
  • Work on multiple features/branches simultaneously

Why They're Perfect for Agents

  • Run multiple agents on different branches
  • No context switching or stashing required
  • Agents can work in parallel without conflicts
  • Test features independently while keeping main clean
$ aem up

Automatic Worktree Detection

aem up automatically detects when it's launched in a Git worktree and will pick a non-conflicting port: run as many dev servers as you have worktrees. Since version 16.12.0 (2025-09-16)

./copresenter "What's your favorite aspect about multi-clauding, my little agentic friend?"

Seeing like an Agent

OBSERVE

    .-"-._.-"-._.-"-._.-"-.
   /                       \
  |   .-----------------.  |
  |   |  .-----------. |   |
  |   | |    * * *   | |   |
  |   | |   /  |  \  | |   |
  |   | |  /___|___\ | |   |
  |   | '-----------'  |   |
  |   '----------------'   |
   \    Agent Vision      /
    '-._.-"-._.-"-._.-"-'

./copresenter "What's that last slide supposed to mean?"

Coding Agents are (mostly) text-based

  • Source code: naturally
  • CLI: very well
  • CLI background tasks: emerging support (Claude is great at that)
  • TUI: early support (gemini), but still buggy
  • Image inputs: mixed: some have it, some don't, but it's always consuming lots of context
  • GUI Apps: no. not yet, at least

To help your agent, see what you see, turn the vision challenge into a coding challenge.

In AEM

$ aem up --forward-browser-logs

Since version 16.13.0 (2025-09-16), aem can forward browser logs to the console, so agents can see them.

Web Development

Use puppeteer or playwright, and instruct your agent to write throw-away scripts to test and capture the page.

./copresenter "In your impartial option, which one is better for web development: puppeteer or playwright?"

Guardrails/Attribution/Transparency

SAFETY

Guardrails/Attribution/Transparency

┌────────────────────────────────────────┐
│                                        │
│                 GitHub                 │
│                                        │
└─────────▲─────────────────────▲────────┘
          │                     │
┌─────────┴───────┐    ┌────────┴────────┐
│    ┏━━━━━━━━┓   │    │    ┏━━━━━━━┓    │
│    ┃        ┃   │    │    ┃       ┃    │
│    ┃   gh   ┃   │    │    ┃  git  ┃    │
│    ┃        ┃   │    │    ┃       ┃    │
│    ┗━━━━━━━━┛   │    │    ┗━━━━━━━┛    │
│  ai-aligned-gh  │    │ ai-aligned-git  │
└─────────────────┘    └─────────────────┘
         ▲                      ▲
         │                      │
         │     ┌─────────┐      │
         │     │ Coding  │      │
         └─────│  Agent  │──────┘
               └─────────┘

Guardrails/Attribution/Transparency

When making changes on github.com (commits, comments, pull requests), attribute them to AI.

This helps reviewers not waste their "Herzblut" on your vibe-coded output.

Agent/Model table

COMPARE
Agent Model
claude claude-opus-4.1
codex gpt-5-high
gemini gemini-2.5-pro
copilot claude-haiku-4.5
cursor-agent composer 1
opencode Grok Code Fast 1
qwen qwen3-coder-plus-2025-09-23
droid Droid Core (GLM 4.6)
amp sonnet-4.5/gpt-5
kimi kimi-k2
crush GLM-4.6
goose gpt-oss-120b 
STOP
DEMO
TIME

./copresenter "How long do you think, did it take to set up these agents?"