David Budáč

Seven decks.
Pick your level.

Same story, from first steps to deep dives. Each deck is self-contained — you don't need the one before it. Find yourself below and start there.

01 Curious, non-technical ——— 02 Developers, new to agents ——— 03 Developers going deeper ——— 04 Devs & IT admins, internals ——— 05–07 Deep dives: cost · orchestration · measurement
01No coding needed The AI Toolbox →

You know what an LLM is. You just don't know which tool to open.

  • You use ChatGPT or Claude and wonder what else exists
  • You don't write code — and don't plan to
  • You want to know what things cost before you ask for a licence

You'll leave with A map of the big three labs, a leader per category (video, voice, avatars, decks, research, automation), five real stories, and how to plug AI into Excel & PowerPoint.

57 slides · no code · prices included

 02For developers · start here Intro to Agentic AI →

You write code and want an agent doing it with you — starting today.

  • You've heard of Claude Code / Copilot but never really used one
  • You want the mental model: model, harness, agent, the loop
  • You want to know what it's allowed to touch, and how to teach it your project

You'll leave with The agent loop, context and why it rots, permissions & safety, memory (CLAUDE.md), skills, plugins and MCP — plus how to work with it well.

43 slides · live demo included

 03For developers · the deep end Agentic Engineering →

You already use agents daily. Now you want to know why they fail — and how to stop it.

  • You want the internals: next-token prediction, attention, statelessness
  • Context engineering is the job — caching, cost, compaction, the dumb zone
  • You're building loops, verifiers, subagents and workflows, not just prompting

You'll leave with Context as the program, token cost & KV-cache economics, model × reasoning-effort routing, the agentic ladder, loops & verifiers (and reward hacking), dynamic workflows.

60 slides · no live demo

 04Deep dive · one topic Subagents & Prompt Caching →

You've done the deep end. Now you want the two primitives that decide what an agent costs.

  • You spawn subagents and want to know exactly what crosses the boundary
  • Your token bill is the problem, not the model
  • You run this on a subscription and need to know what's allowed

You'll leave with Subagent isolation and the four context crossings, parallel fan-out and its limits, cache breakpoints, exact-match prefixes, TTL & eviction, break-even math — and the subscription-vs-API-key rules.

39 slides · devs & IT admins · 30 min

 05Deep dive · level upWIP Cost & Context →

You already use Claude Code every day. Now find out what it actually bills — and stop wasting the window.

  • You want the real numbers — tokens, subscriptions, API, Bedrock
  • You've never traced what one session costs, line by line
  • Your CLAUDE.md has grown and nobody's measured it

You'll leave with Three pricing meters, a which-plan-for-which-team table, the context lifecycle, CLAUDE.md hygiene, /clear vs /compact — and a live bloated-vs-lean demo.

28 slides · experienced devs · ~30 min

 06Deep dive · level upWIP Orchestrating Agents →

One agent is table stakes. Run several — safely — and give them the right capabilities.

  • You're juggling agent terminals and want a dashboard, worktrees, or a backlog-driven fleet
  • You want sandboxing before you let agents run unattended
  • You keep debating skills vs CLAUDE.md vs MCP vs CLI

You'll leave with A map of the orchestration ecosystem, the three failure modes of parallel agents, skills & plugins for teams, and a per-capability MCP-vs-CLI heuristic.

32 slides · experienced devs · ~30 min

 07Deep dive · level upWIP Measuring What Works →

Stop arguing about prompts from vibes. Measure — then route models and harnesses by the numbers.

  • Your team tweaks CLAUDE.md and prompts with no way to compare
  • You want to read traces for wasted turns and real cost
  • You're curious about open-weight models, local inference & alternative harnesses

You'll leave with A minimal eval loop you can start this week, trace-reading skills, the July-2026 open-model landscape (Gemma 4, Qwen 3.6, GLM-5.2), and daily habits that compound.

31 slides · experienced devs · ~30 min

 08Conference talk · case study Has this happened before? →

Oracle logs tell you what failed. They don't tell you whether you've seen it before, or what fixed it last time.

  • You run Oracle databases and get the same ORA errors again and again
  • You want a wiki of past incidents that builds itself from the logs
  • You work somewhere that can't ship its logs to a cloud agent

You'll leave with Logbook: an agent that turns Oracle logs into a git-backed incident wiki, with a worked ORA-04031 case study, the local/cloud boundary and which model gets which job.

42 slides + 35 backup · ~29 min · AgentConf

 09Meetup · agent workflows T3 Code →

Coordinating agents across harnesses and machines.

  • You run several coding agents and need to keep track of them
  • You want to delegate across harnesses and projects
  • You want to run work on another machine and inspect the results

Covered in the talk The thread inbox, delegation, remote machines, previews, PRs and automation in T3 Code.

20 min · 3 November 2026 · David Budac

Still not sure? One question each

Do you write code?
No → The AI Toolbox
Has an agent ever edited your repo?
Not yet → Intro to Agentic AI
Do you know why a long chat gets dumber?
No, and I want to → Agentic Engineering
Do you know why a cache hit is 10× cheaper?
Not exactly → Subagents & Prompt Caching
Know what your agent sessions actually bill?
Not really → Cost & Context WIP
Ever run more than one agent at once?
Want to → Orchestrating Agents WIP
Still judging prompt changes by feel?
Yes → Measuring What Works WIP