"I spent two days at AI Native DevCon. Most of the noise around AI is mysticism. Three themes cut through it in a way that's directly relevant to how we build."
1
Act 1
What AI actually is, stripped of hype
2
Act 2
The new engineering stack it's producing
3
Act 3
How your team works inside that stack
Act 1
What AI Actually Is
"A new kind of software that is good at pattern recognition"
Not magic. Not (yet) thinking. The core idea is 70 years old — what changed was compute + data, not a new algorithm. Every time someone says "the AI thinks X", replace "thinks" with "statistically continues a pattern". It changes how you evaluate risk.
Act 1
The Two Practical Hazards
Training-Data Bias Is Invisible
A wolf vs. husky classifier was actually detecting snow in the background. A skin cancer app was detecting doctors' rulers. Silent failures: no error thrown, confident output, wrong answer.
The Black Box
Nobody knows why a trained network works — not even the people who built it. It works until it doesn't. Maintenance is not optional.
Act 2
A Recognisable New Stack Is Forming
"This stack is already here. Claude Code is a harness. CLAUDE.md is context. The question isn't whether we're building on this stack — it's whether we're doing it deliberately."
1
Models
New compute primitives — the "operating systems"
2
Tools
CLI, MCP, APIs: deterministic arms and legs
3
Context
Everything that programs the model (rules, skills, passive context)
4
Harnesses
Deterministic software wrapping the probabilistic model
5
Factory Lines
Composed harnesses: repeatable pipelines
Act 2
Context Engineering
If code programs the machine, context engineering programs the model.
Anthropic's Applied AI team defines it as the primary lever for turning raw model intelligence into "durable, scalable, useful product" — and "a really great investment... [that] has the effect of multiplying the intelligence even as models get smarter."
Context comes in three forms, each solving a different problem:
Skills use progressive disclosure: the agent scans a short frontmatter to decide relevance, then loads the full body only when needed. Think of it as a bookshelf — you scan titles, pull out the book you need.
Context engineering is the discipline of designing all three layers intentionally — what goes in, when it loads, and how it composes. If a skill is wrong, the agent confidently does the wrong thing with perfectly good code underneath.
Act 2
Context Engineering Needs the Same Disciplines as Code
~2M
Skills on GitHub
Up from ~0 at the start of 2025
30%+
Were Malicious
In one open-source ecosystem alone
Act 3
How Your Team Works in This Stack
"The Humans Architect the System, the AI Writes the Code"
"Vibes don't scale"
Before
Engineers write implementation
Now
Engineers own intent, architecture, and constraints; agents implement
The value moves upstream — into the decisions the agent cannot make.
Act 3
Intent Must Be Explicit and Durable
Behaviour acceptance criteria defined upfront are the source of truth — not a passing CI run, not the agent's confidence.
1
Intent Document
What the system must do
2
Constraint List
What must not change; what patterns must be preserved
3
UAT Criteria
The test that proves the output is correct
What This Means for Our Team
1
Treat Skills as Code
Version them, scan them, eval them, own their lifecycle.
2
Write Intent First
Every agent session starts with a constraint list and UAT criteria — not a vague prompt.
3
Watch for Silent Failures
Confident AI output is not correct output. Review gates must check against original intent.