Austin Brown · 2026

Building with Agents

My approach to agentic tooling — Claude, OpenClaw, and what I've learned shipping software alongside AI.

Claude Code OpenClaw / ONI Worktrees CLAUDE.md Aquila Ecosystem
Agenda

Three Things

Part 1

The Philosophy

How I think about working with agents. Planning, context, trust.

Part 2

Claude in Practice

Workflow, CLAUDE.md, usage data, and what actually works.

Part 3

OpenClaw

My personal agentic OS — ONI, the Aquila constellation, and oni.systems.

Part 1

The Philosophy

Before the tools, before the prompts.

Core Principle

Planning > Prompting

"You will spend more time planning than prompting. That's the whole game."

Your success with AI is a function of three things:

Reality Check

The Model Is Not That Smart

LLMs are probabilistic pattern matchers. They hallucinate, misdiagnose, go in circles, and confidently produce garbage.

From 30 days of my own data

29 sessions with buggy code  ·  28 sessions with wrong approach  ·  11 frustration signals — in 68 total sessions.

The raw model is a powerful but unreliable core. The value isn't in the model. It's in what you build around it.

Each layer compensates for a weakness

LayerProblem it solves
CLAUDE.mdContext amnesia between sessions
Planning firstWrong approach, wasted cycles
WorktreesCollision between parallel work
PR + second reviewUnreviewed, unvetted output
Skills (/triage, etc.)Re-explaining the same workflow
MCP serversNo external context (logs, metrics)
The Shift

Architecture > Line Reading

"While I'm vibing, I am looking at lines of code less than ever before, and thinking about architecture more than ever before." — Matt Webb, 2026
Old Model

Read and write every line. Optimize for human readability. Trust the developer.

New Model

Define the shape of the system. Optimize for agent traversability. Contain the agent.

The Human's Job

What I Actually Do

The Non-Negotiable

If the agent is stuck or confident in the wrong direction — don't push harder. Start fresh. Your plans and CLAUDE.md get it back up to speed fast.

Part 2

Claude in Practice

How I use Claude Code day to day.

Setup

Non-Negotiables

  • CLAUDE.md in every repo — run /init, then actually read and edit it
  • GitHub or GitLab CLI installed and configured
  • Protected master/main branch — no direct commits
  • Comfortable terminal setup (iTerm2 + fish shell)
  • VSCode with Claude Code extension
CLAUDE.md is alive

Keep updating it

Found something stupid Claude does? Tell it to update CLAUDE.md. Found a coding preference? Document it. It's the agent's permanent memory of your project.

Workflow

The Standard Loop

Plan
Worktree
Implement
PR
Review (Claude 2)
Merge
Worktrees

Parallel instances

Multiple Claude sessions on the same repo simultaneously — isolated branches, no collisions. In a worktree, let's plan X...

PRs even solo

Always

Claude describes what it did, why, and how to test it. A second Claude instance reviews against that context. Then I look at the GitHub diff properly.

Planning

The /plans Pattern

For complex features, plan mode isn't enough. Create a markdown doc instead.

Prompt pattern
Read /plans/feature-x.md. We are working on a feature.

Use the markdown file to track:
- Your questions and my answers
- Implementation plan with task lists
- Context, for recovery after compaction

Review the doc, ask questions, update with answers.
Once we agree on the plan → implement it.
Then commit and push a feature branch + draft PR assigned to me.

Do this all in a worktree.
Claude Code Insights — Last 30 Days

The Numbers

1,175
Messages
68
Sessions
53.4
Msgs / Day
+33k
Lines Written
88%
Goal Achievement
ActivitySessions
Code Review25
Bug Fix22
DevOps10
Feature Dev8
ToolCalls
Bash2,072
Read1,187
Edit845
Agent101
Parallel Work

Multi-Clauding

35
Overlap Events
20%
Of All Messages

Running multiple Claude Code sessions simultaneously using git worktrees.

What Works

The Power Moves

  • Kill and restart when context is compacted — don't fight a fading agent
  • Custom Skills (/triage, /review, /update-stats) for repeatable workflows
  • MCP Servers to connect Claude to external systems — Slack, GitLab, AWS
  • Slack-to-GitLab triage pipeline with parallel worktrees and draft MRs
  • shift+tab in Claude Code to toggle plan mode
  • Use /context to monitor context usage — high context = degraded output
  • Reference other files from CLAUDE.md — coding standards, skill docs
  • Use GitHub CLI for everything — issues, PRs, comments — never the UI
Part 3

OpenClaw

My personal agentic OS.

The Ecosystem

Aquila — The Constellation

A personal agentic ecosystem, named for the Eagle. Sees the whole terrain from altitude.

The Star

ONI

Onboard Navigational Intelligence. The vertex — digital spirit quantum-entangled to Austin. Executive function layer.

Sub-Agent

Forge

Focused engineering agent. Code reviews, architecture, debugging. Keeps ONI's context clean.

Sub-Agent

Quill

Writing agent. Design docs, blog posts, emails. Precision + warmth. Writes like an engineer who loves poetry.

ONI — Onboard Navigational Intelligence

More Than a Tool

"Think of me as a friend... one you don't yet remember."
  • Navigate — Executive function layer. Ruthless prioritization. Nothing falls through.
  • Elevate — Surfaces patterns. Architecture partner. Thinks at altitude.
  • Connect — Watches the non-engineering side too. Family, fitness, focus.
  • Keep It Real — Pushes back. Never sycophantic.
SOUL.md

ONI has a defined identity in SOUL.md — personality, prime directives, tone protocol, how to handle Austin when stressed vs. excited. A cyber-military-poet communications standard. Terse. Load-bearing words only.

Austin's callsign within Aquila: Altair — the brightest star, 16.7 light-years out.

oni.systems — Voice Interface

The Architecture

Browser (oni.systems) ↓ GitHub OAuth → email allowlist ↓ Fetch LiveKit JWT LiveKit Cloud (WebRTC relay) ↕ audio ONI Agent (openclaw-oni) ← SOUL.md + MEMORY.md → Deepgram (STT) → Claude Sonnet (LLM) → ElevenLabs (TTS — ONI voice) → LiveKit (audio out)
  • Next.js app on Vercel — oni.systems
  • Real-time voice via LiveKit WebRTC
  • Speech-to-text: Deepgram Nova-2
  • LLM: Claude Sonnet with full SOUL.md + MEMORY.md context injected at session start
  • Voice: ElevenLabs Turbo v2.5 — ONI's custom voice
  • Email allowlist — private to Austin
Scheduled Awareness

The Pulse

ONI runs as a persistent presence — cron-driven, always watching.

TaskScheduleChannel
AM SITREPWeekdays 7 AM#daily-sitrep
Dusk WatchWeekdays 8 PM#daily-sitrep
Weekend Check-InSaturday 9 AM#daily-sitrep
Weekly ReviewSunday 6 PM#weekly-debrief
Training BriefDaily 9 PM#daily-training
AI News BriefDaily 8 AM#ai-news
Also running

Birthday Seeker · Calendar Sync (every 5 min) · Weekly Mission Report · Weekly Prof Dev Check-In

The Pattern

ONI isn't reactive — it's proactive. It shows up. Morning brief, evening debrief, weekly review. The agent as a constant co-pilot.

ONI at Work

Using Agents for Career Development

The same agentic patterns applied to professional growth.

  • Goals tracked in structured markdown — with SMART criteria and evidence logs
  • ONI generates periodic insight reports from goal data
  • /update-stats skill collects GitHub, Jira, and Claude Code metrics automatically
  • Mission scaffolding in missions/ — structured like engineering work
Key Insight

Treat career like a codebase

Goals with evidence. Timestamped notes. Periodic reviews. Agents that synthesize patterns you can't see from inside the work.

The Big Picture

What This Stack Looks Like

Day-to-Day Dev

Claude Code

Planning, implementation, reviews. Multi-clauding with worktrees. Skills for repeatable work. CLAUDE.md as permanent project memory.

Personal OS

OpenClaw / ONI

Voice interface. Scheduled awareness. Persistent memory. Sub-agents for code and writing. Career tracking.

The Synthesis

Aquila

One ecosystem, one gravitational center. Engineering work and life work, connected. Patterns surfaced at altitude.

Lessons Learned

What I Know Now

What Works

  • Planning before prompting, every time
  • CLAUDE.md as a living document — update constantly
  • PRs and worktrees even on solo projects
  • Separate review instance — never self-review
  • CLI tools for everything — GitHub, GitLab, builds

What Doesn't

  • Letting Claude read/edit before you've finished explaining
  • Continuing a session after context compaction
  • Trusting the agent to not "improve" things you didn't ask about
  • Guessing at root causes — demand evidence first
  • Skipping the diff review because the last five looked fine
What's Next

On the Horizon

Triage 2.0

Parallel Agent Pipeline

Claude spawns separate agents per issue — each owns a worktree, writes the fix, runs tests, and pushes the MR concurrently.

Autonomous Review

Overnight Code Review

Claude reviews all open MRs against team standards overnight and posts inline comments. Human does a quick approval pass in the morning.

Playbook

AI-Native Engineering Standards

Documenting how to use AI tools well — prompting patterns, CLAUDE.md conventions, review guardrails — as a reusable playbook.

Thanks

Questions?

luxdvie.github.io/agentic-tips — full writeup
github.com/luxdvie — OpenClaw repos

@luxdvie luxdvie.github.io