From Sidebar to Sub-Agents
A Live, Hands-On Intro to Building with AI Agents
Kyle Hoehns · Source Allies
About Your Speaker
Kyle Hoehns
Staff Software Engineering Consultant · Source Allies
- 15+ years shipping complex, production software.
- Backend roots in Java & Go — scalable, efficient systems across many industries.
- Firm believer in not taking myself too seriously.
LIVE CODING AHEAD
First, check the weather
Level Up
LEVEL 0 / 6
We start from nothing
Blank repo. One trivial endpoint. Zero AI.
Demo: this is our baseline.
LEVEL 1 / 6 · GITHUB COPILOT · HAIKU
AI as a smarter search box
Ask a question, paste a snippet.
Watch: it answers, but your hands still do all the work.
LEVEL 2 / 6 · GITHUB COPILOT · HAIKU
Hands move from your keyboard to the assistant's
Describe the intent; it makes the edit.
Watch: you stop typing code and start directing it.
LEVEL 3 / 6 · SWITCH TO CLAUDE CODE · SONNET
Give it the terminal
It adds persistence, runs the build, and fixes its own errors.
Watch: the self-correction loop. (Gotcha next.)
Gotcha: the Maven reflex
- What bit us: on a Gradle project, the agent kept reaching for Maven.
- The fix: an
AGENTS.mdwith project context. - Lesson: give agents an
AGENTS.mdwith what they need every interaction — build & test commands, how you write tests, version gotchas (“don't use Java 21 features”). Capture what they get wrong, not niche trivia. - Also: the standards are still settling — tooling isn't uniform yet.
LEVEL 4 / 6 · BACK TO HAIKU · ON PURPOSE
Teach it your standards once
A custom skill auto-triggers and applies your conventions.
Watch: it picks the right imports without being told.
Gotcha: stale training vs. a breaking change
- What bit us: Spring Boot 4 moved Jackson to
tools.jacksonand relocated@AutoConfigureMockMvc. A fast model confidently used the old imports. - The fix: the skill carries the current facts, closing the gap.
- Lesson: skills inject knowledge the model's training doesn't have.
- Bonus: with the stack facts in the skill, cheap Haiku produced the same correct tests as Sonnet. Routing down is viable when context is good.
LEVEL 5 / 6 · SONNET
Let the build hold the line
Quality gates — coverage and formatting — fail the build until the agent complies.
Watch: the same change goes red, and the agent writes the tests itself.
Guardrails turn intent into enforcement
Without guardrails
- Ask for a feature — it ships the feature, nothing more.
- No tests, no formatting pass — nothing told it to.
- The agent isn't lazy; it just can't see your standards.
With guardrails
- Same prompt —
./gradlew buildgoes red on the 80% coverage gate. - It reads the failure and writes the tests to get green.
- Coverage, formatting, lint — the build enforces them, not you.
A failing build is the clearest spec you can hand an agent — enforce it, don't hope for it.
From doing to delegating
No vibe coding: you're in the driver's seat
Spec it, scope it, define done — that's what makes it safe to hand off.
LEVEL 6 / 6 · SONNET
Your first direct reports
A skill orchestrates a team to build a feature — a four-wide review + docs fan-out.
Watch: the parent delegates, collects, and applies fixes.
Zooming out
It all comes down to the context window
Managing this window is the real skill now — “context engineering.” Sub-agents are firewalls: each floods its own window with noise, hands back only signal.
Skill or agent?
Skill = a recipe
- Loads into your context window.
- Teaches HOW — a procedure, your conventions, fresh facts.
- Uses the tools you already have.
- Best for: reusable expertise any agent should apply.
Agent = a coworker
- Spawned as a sub-agent — its own context window, tools, and model.
- Isolates WORK — noise, big or parallel tasks.
- Scoped tools — least privilege by role.
- Best for: work that floods your context or fans out.
A skill changes behavior; a sub-agent protects context — and they compose. Stack them and one command can run your whole pipeline.
Kick it off, go to lunch
- ★ You leadplan it, then kick it off
- read the ticketJira or GitHub
- planADRs & a checked plan
- ↓ “ship it” — your last keystroke
- 🤖 Agent runs solothe rest is automatic — go to lunch 🍔
-
build the featurein a worktree, a team fans out:
developer tester reviewer ×3 doc-writer 🔒 tests · coverage · format
- shipcommit · push · open the PR
- adversarial review ↻PR bots → auto-address → re-run
- notify youback from lunch to a green PR
Generation is cheap. Review is the bottleneck.
- The flip: writing code is nearly free now — the scarce thing is trusting what comes back.
- The evidence: GitHub has run 60M+ Copilot code reviews; AI-heavy teams merge far more code, but review time balloons — even METR couldn't measure a clean speedup in 2026.
- The answer:
ship-featurespends most of its steps on review — the gates, three scoped reviewers, and the bots on the PR.
Your job moves up the stack: delegate, review, own.
An empty project became a team of agents — and you know how to start.
Start small: pick one bounded task, let the assistant draft it, you review.
Thanks for hanging out!
