Let's Talk Agentic: Field Notes

Martin Povolny
Red Hat — Cost Management

Agenda

1. slack-search — separate deck
2. dictpert — separate deck
3. Personal Productivity — wiki, MarkdownViewer, Claude Code, open-code
4. In-browser AI (optional)

Personal Productivity

Wiki · MarkdownViewer · Claude Code · open-code

The Workflow Shift

Engineering is moving to a higher level of abstraction.

Before: code → format → write docs → edit HTML → code
The "muscle memory" tasks gave your brain rest periods.

Now: review → design → review → decide → review → decide
Every task is high-level and mentally intensive. No breaks.

We read and review now where we used to write.
The cognitive load of engineering work is increasing — a new kind of fatigue.
A new equilibrium will have to be found. Companies and people are figuring it out right now.

The Problem with Knowledge Bases

Previously used: macOS Notes, Joplin, Craft, OneNote, ...

Knowledge bases rot.
The tool changes. The computer changes. The job changes.

The knowledge base needs to outlast all of them.

LLM Wiki — the Karpathy Pattern

• Human curates sources and asks questions
• LLM writes and maintains all wiki content
• Plain markdown + Obsidian — no vendor lock-in
• Git versioned — survives tool changes, job changes, everything

"The LLM solves the maintenance problem that killed every personal wiki before it."

What's in the Wiki

wiki/                   77 sources, 250+ pages
├── raw/          ← immutable sources (human adds)
├── sources/      ← LLM-written summaries
├── concepts/     ← topic pages (cheatsheets, glossaries)
├── entities/     ← people, tools, organizations
├── projects/     ← one page per project
├── tasks/        ← action items with status tracking
├── daily/        ← daily notes (raw → processed by LLM)
└── meetings/
    └── auto/     ← Gemini auto-generated notes

Custom Commands

• Obsidian Web Clipper for quick source capture

Meeting Notes Pipeline

Google Meet (Gemini) → Email → Apps Script → Sheet
Google Meet recording → Drive  → Apps Script → Sheet
                                                 ↓
              /process-meetings → wiki/meetings/auto/

• Automated via Google Apps Script (hourly)
• Cross-links auto notes manual notes
• Extracts recording URLs from Google Doc rich links

MarkdownViewer

The agentic era produces a lot of markdown — agent output, design docs, PRDs, meeting notes, wiki pages. We need a good flow for reading, editing, and producing documents.

MarkdownViewer — macOS native app, vibe-coded by Ladas, Libor, Martin

github.com/Ladas/MarkdownViewer

MarkdownViewer Features

• Custom themes — branded per project or company
• Tabs — multiple documents open at once
• Export to Google Docs — one click
• Claude-assisted editing — AI help inside the viewer

MarkdownViewer — Themes in Action

Blue · Red Hat · Koku
Custom CSS per project
click to enlarge

Claude Code Workflows

The coding agent as a personal operating system

Background Agents

claude agents — like a built-in tmux for AI sessions.
Automatically tracks all running agents, attach/detach anytime.

You ──→ Agent 1: "explore the codebase for X"
   └──→ Agent 2: "review this PR"
   └──→ Agent 3: "run the test suite"

Spawn agents for parallel work — results flow back without blocking.

/loop and /goal

Use with care:

• Active development — behavior changes between releases
• "Split brain" — context compaction during /loop or /goal can cause the agent to lose its thread

Custom Commands, Skills & MCP

Custom slash commands = markdown instruction files in .claude/commands/

• /ingest, /lint, /process-meetings, /publish — all commands in this wiki
• Simple to write, project-scoped, version-controlled with the repo

Skills = more complex reusable packages (plugin marketplace)

• /slack-search — installed as a plugin, cross-project
• Progressive disclosure: agent loads full instructions only when task matches

MCP = tool servers the agent can call

• Google Docs/Drive/Sheets/Calendar (meeting notes, create docs)
• Jira (create tickets, search issues)
• Playwright (browser automation, UI testing)

open-code

Open-source terminal coding agent (MIT, 176k+ stars, 75+ LLM providers)

• Free BYOK — bring your own API key or local inference (e.g. LM Studio, Ollama)
• "Go" subscription ($5→$10/mo) — curated open-weight models, zero setup
  • GLM 5.2, Kimi K2.7, DeepSeek V4, Qwen 3.7, MiniMax M3, ...

Why OSS Models Catch Up

GLM 5.2 — exceptional coding benchmarks from an open-source model

How? The training recipe is now known:

1. Take a base model, introduce subtle bugs into real codebases
2. Have Codex/Claude Code fix the bugs — record all inputs/outputs
3. Use those traces for supervised finetuning (conversational problem-solving)
4. Use test script pass/fail for reinforcement learning (actual problem-solving)

The student becomes stronger than the teacher — verifiable results + solution space exploration.

"Coding LLMs is a solved problem and everyone will end up having the same solution."
— Burkov

Thank You

Martin Povolny

Questions?