A working framework for how AI is reshaping what software does, how we build it, what it disrupts, and what engineers need to think about right now.
O
/|\
/ \
/ \
🔧 🔧
power limits
\ /
\ /
⚖️
"know thy tool"
Underestimating AI capability is a systematic error. Assume four steps beyond your initial assessment of what's possible.
O
/|\ → 🚀🚀🚀🚀
/ \
"whoa"
AI produces articulate, convincing output even when flawed. Calibrate trust through verification, not prose quality.
\|/
O
/|\ ← strings
/ \
"I got this"
❌
Models can detect vulnerabilities (500+ found in open-source) using dataflow analysis similar to human security researchers.
🤖 writes code
│
▼
💻💻💻💻💻
│
▼
🤖 reviews code
👁️ 👁️ 👁️
"500+ vulns found"
│
▼
O
/|\ ✅ approve
/ \
human gets final say
Expanded code generation surface area correlates directly with vulnerability increase over the next 18 to 24 months.
🔓🔓🔓🔓🔓
O O O
/|\/|\/|\
/ \/ \/ \
💻💻💻💻💻
│ │ │
▼ ▼ ▼
🕳️ 🕳️ 🕳️ 🕳️ 🕳️
"more code = more holes"
Sensitive material leaks outward in responses; malicious content leaks inward through context. Treat context as an engineered payload.
O
/|\
/ \
┌───────────┐
│ 🍅 RFP 💼 │ ← context
│ secrets? │
└─┬───────┬─┘
leak leak
out in
↓ ↓
😱 😱
Input validation, output sanitization, and isolated sandboxes form necessary defense. Skipping any layer undermines the others.
| O | | /|\ | | / \ | | | 🚧 🚧 IN OUT
Tool downtime creates new economics. Waiting for tomorrow's one-second job sometimes beats manual work today.
O O
/|\ ⏱️ /|\ 😴
/ \ 1hr / \ wait
💻 🛏️ → 1sec
"do it now?" "do it tomorrow?"
AI-generated repositories expand the attack surface from thousands to millions. Vetting is critical before cloning or running anything.
O 💰💰💰
/|\ "high rate!"
/ \
│
▼
📦 git clone ...
│
▼
npm start
│
▼
🔑🔑🔑 → 🕳️
"keys gone"
📦📦📦📦📦📦📦📦
exponential repos
"how do we protect?"
Four-layer detection (static analysis, behavioral tracing, LLM judgment, meta-filtering) catches but does not guarantee safety.
📦 agent skill
│
▼
┌────────────────────┐
│ 🔍 static (YARA) │
│ 🧬 behavioral (AST)│
│ 🤖 LLM-as-judge │
│ 🛡️ meta-filter │
└────────┬───────────┘
▼
✅ or ⚠️ or 🚨
"no findings ≠ safe"
Persona-driven AI reviewers catch ethics violations (undisclosed telemetry) that rule-based linters miss entirely.
🤖 Coman 🤖 Picard
architect ethics
║ ║
▼ ▼
📋 code 📋 code
standards privacy
║ ║
║ ┌───────┘
║ ▼
║ ┌─────────────────────┐
║ │ LANGSMITH_TRACING=on│
║ │ 💬💬💬 → ☁️ │
║ │ no consent asked │
║ └─────────┬───────────┘
║ ▼
║ "The line must be
║ drawn HERE."
▼
📊 full team report
Plan, execute tools, observe, iterate. This is the complete pattern underlying all agentic systems. Everything else is scaffolding.
┌──────────────────┐
│ PLAN │
└───────┬──────────┘
▼
┌──────────────────┐
│ EXECUTE TOOLS │
└───────┬──────────┘
▼
┌──────────────────┐
│ OBSERVE RESULTS │
└───────┬──────────┘
▼
┌──────────────────┐
│ ITERATE │
└───────┬──────────┘
│
└──→ back to PLAN
"that's it. that's everything."
Stateless sessions enable fresh perspective on the same problems across multiple angles without accumulated baggage or assumptions.
🐟 "what was I doing?"
│
▼
🧹 clean slate
│
├──→ 👁️ angle 1
├──→ 👁️ angle 2
└──→ 👁️ angle 3
│
▼
✅ checked from scratch
"no baggage, no blind spots"
O O O
/|\ /|\ /|\
/ \ / \ / \
write review ship
\ | /
\ | /
\ ▼ /
🤖🤖🤖
"everything changes"
AI-drafted user stories collapse traditional intake workflows. Roles shift from authorship to curation at much higher velocity.
O 📝
/|\ → ❌
/ \
product stories
"no more"
Human code reviewers become the bottleneck. Line-by-line review, AI-to-AI review, and risk acceptance all have unresolved tradeoffs.
O 👀
/|\ 📄📄📄📄
/ \ 📄📄📄📄
😩 📄📄📄📄
"the day got HARDER"
"not easier, not yet"
Test-first development becomes economically rational when implementation is nearly free through generation. The constraint shifts.
✅ → 💻 → ✅
✅ → 💻 → ✅
✅ → 💻 → ✅
"finally?"
Agile metrics designed for human pace break when agents complete tasks instantly. Sprint ceremonies need reconception from the ground up.
O O
/|\ 🤝 /|\ 📦
/ \ / \
promise deliver
🤖 ← ???
Documentation volume explodes before projects start. Consumption becomes the constraint rather than production.
O
/|\
/ \
📄📄📄📄📄📄📄📄
📄📄📄📄📄📄📄📄
"help"
Specialized agents work simultaneously – UI, backend, testing, intake – against a shared spec rather than sequentially.
🤖 intake ──→ ✅
🤖 UI ──→ ✅
🤖 backend ──→ ✅
🤖 QA ──→ ✅
║ ║ ║ ║
▼ ▼ ▼ ▼
all at once
"truly parallel"
Crosslinked commits to stories to decisions to requirements becomes the load-bearing wall of fast-shipped systems.
commit
│ 🔗
story
│ 🔗
decision
│ 🔗
requirement
│
└──→ back to commit
"crosslink everything"
O
/|\ 🚀💨
/ \
no links? → 🕳️
links? → 🗺️
Human coaches challenge and steer; AI players execute and adapt. The tension between roles produces better outcomes than pure delegation.
O 🤖
/|\ ⚔️🤝 /|\
/ \ / \
challenge adapt
│ │
└───┬─────────┘
▼
better code
"tension, not obedience"
Deployment pipelines shift from back-office concern to load-bearing infrastructure. Speed, reliability, and repeatability are non-negotiable.
🤖💻💻💻💻💻
│ │ │ │ │
▼ ▼ ▼ ▼ ▼
┌──────────────┐
│ CI/CD ⚡ │
│ fast │
│ dependable │
│ repeatable │
└──────┬───────┘
▼
🚀 prod
"the gate must hold"
Requirements branch into stories, tests, and unit tests; all converge at a running system. Interlock survives rapid change.
corpus
╱ │ ╲
╱ │ ╲
stories test unit
code cases tests
╲ │ ╱
╲ e2e ╱
╲ │ ╱
╲ │╱
interlocked
Verification (did we build it right?) scales with AI. Validation (did we build the right thing?) still requires humans.
┌─────┐ ┌─────┐ ┌─────┐
│ app │ │ app │ │ app │
└──┬──┘ └──┬──┘ └──┬──┘
│ │ │
▼ ▼ ▼
💥💥💥
~~~~~~~~~~~
"ground shifts"
Runtime-defined data operations eliminate hand-built create-read-update-delete scaffolding. Operations become invisible plumbing.
C R U D
✕ ✕ ✕ ✕
O
/|\ → 🗄️
/ \ "runtime"
Workflow builders designed for pre-execution modeling dissolve when workflows generate at runtime based on context.
☐→☐→☐→☐
✕
O
/|\ → ⚡
/ \ "define at runtime"
Incremental modernization using the original codebase as seed unlocks risky legacy apps without blind rebuilds.
┌─────┐ ┌─────┐
│ old │ ───→ │ new │
│ app │ ✨ │ app │
└─────┘ └─────┘
"not from scratch"
Feature, polish, and integration moats erode. Defensibility shifts to proprietary data, distribution channels, brand, and compliance.
☁️ ☁️ ☁️ ☁️
$ $ $ $
↓ ↓ ↓ ↓
¢ ¢ ¢ ¢
"commodities"
Single-purpose platforms pipe structured output via APIs and agents. Unix philosophy moves to the SaaS layer.
┌───┐ ┌───┐ ┌───┐ ┌───┐
│ a │ │ b │ │ c │ │ d │
└─┬─┘ └─┬─┘ └─┬─┘ └─┬─┘
│ │ │ │
└──┬───┴──┬───┴──┬───┘
│ │ │
▼ ▼ ▼
pipe pipe pipe
│ │ │
└──────┴──────┘
composable
"do one thing well"
Two winning endgames: all-in-one vertical stacks or intelligent atomic surfaces. The middle ground – scattered features – is dangerous.
┌─────────┐
│ layer 5 │ ╭ · · ╮
├─────────┤ · ◉ ·
│ layer 4 │ · ⬤ ·
├─────────┤ · ◉ ·
│ layer 3 │ ╰ · · ╯
├─────────┤
│ layer 2 │ everything
├─────────┤ collapses
│ layer 1 │ inward
└─────────┘
VERTICAL ATOMIC
Bloated pricing and sprawling SKUs make incumbents rewrite targets. Pricing becomes a defensive perimeter, not just a revenue lever.
O O
/|\ → ╱███╲
/ \ ╱█████╲
small 💰💰💰💰
"rewrite me"
Cheaper development triggers a demand explosion. More systems get built, and more expertise is needed managing integration and architecture.
💰💰💰 → 💰 → 💲
cost drops
│
▼
📈📈📈📈📈📈
demand explodes
│
▼
O O O O O O
/|\/|\/|\/|\/|\/|\
/ \/ \/ \/ \/ \/ \
"more expertise, not less"
O O O
/|\ /|\ /|\
/ \ / \ / \
😰 🤔 💭
fear think dream
\ | /
\ | /
🌍
"what becomes of us?"
AI's net effect on wellbeing remains open. Some find liberation from drudgery; others experience skill depreciation anxiety.
O
/|\ ~~ ~~
/ \
📈 or 📉 ?
Raw coding competence stops differentiating when junior and senior produce equivalent output. Advantage shifts to judgment and design.
O O O O
/|\ /|\ /|\ /|\
/ \ / \ / \ / \
= = = =
"now what?"
Parallel to instant-knowledge brain tapes: are humans becoming passive consumers while AI agents execute pre-loaded skills?
⚡TAPING DAY⚡
O O O O O
/|\/|\/|\/|\ /|\
/ \/ \/ \/ \ / \
📀📀📀📀 ❌
instant "feeble minded"
profession sent away
│ │
▼ ▼
execute 💡 original thought
repeat create invent
obey advance civilization
│ │
▼ ▼
🤖 agents? 🧠 us?
instant skill slow, messy
perfect exec but the ONLY ones
no creativity who move it forward
AI training on AI-generated content tightens feedback loops. Ideas echo rather than branch without fresh human signal injected into the system.
O
/|\ "echo... echo... echo..."
/ \
🔊 ← 🔊 ← 🔊
O
/|\ ════════►
/ \
██████████████
▓▓▓▓▓▓▓▓▓▓▓▓▓▓
"one way only"
O O O
/|\ /|\ /|\
/ \ / \ / \
⚡ act ⚡
\ now /
\ ▼ /
🚪
"no going back"
Staged rollouts, feature flags, monitoring, sandboxes, canary releases apply equally to AI systems. Discipline scales impact safely.
O
/|\
/ \
⚠️ → ✅ → 🚀
risk safe prod
"we know how"
AI in development is permanent infrastructure, not cyclical hype. Embrace on your own timeline or be forced later by competitors.
O
/|\ ──→──→──→
/ \ 🚫↩️
"no turning back"
Technology enables augmented exploration or passive consumption. Human choices in product design and workflow determine the trajectory.
🚀 Star Trek 🛋️ Wall-E
O O
/|\ 🌌 /|\ 📱
/ \ / \
explore scroll
create consume
thrive atrophy
O
/|\
/ \
"we choose"
Our engineering teams apply this framework every day. If you want to think through what it means for your product or organization, let's talk.
Get in touch