Report

Course-to-Project Scan

Welche Teile des Thinking-in-Agents-Kurses bereits in eigenen Projekten sichtbar sind — inklusive Workshop-Beispielen.

Scan date: 2026-06-15 · Source: D:/course/chapters, D:/projects, C:/code, C:/dev, D:/workspaces/workshopper

Date: 2026-06-15

Scope scanned

Course source reviewed:

  • D:/course/chapters/01_foundations through 10_vocabulary
  • D:/course/chapters/new_agent_os/agentic_os.md
  • workshop workspace: D:/workspaces/workshopper

Project locations scanned at top level, then sampled by README / AGENTS / CLAUDE / package metadata:

  • D:/projects
  • C:/code
  • C:/dev
  • D:/workspaces/workshopper

Executive summary

You are already using a large part of the course material in your own projects. The strongest overlap is not generic “AI app” usage; it is the practical infrastructure around agentic engineering:

  1. Harness thinking — torbjorn, symphony, pi_action_ui, pi_auto_runner, handsfree-pi.
  2. Context as architecture — C:/dev/AGENTS.md, C:/dev/_brain, C:/dev/_skills, D:/workspaces/workshopper, project AGENTS.md / CLAUDE.md files.
  3. Backpressure and quality gates — torbjorn, C:/dev/_gate, D:/projects/tlslide, D:/projects/reed.
  4. Agent-facing tools — gchatctl, jiractl, porkctl, brainctl, notifyctl, xpostctl, sapctl, etc.
  5. Workshop material already exists — D:/workspaces/workshopper/workshops/vibe-coding-to-agentic-engineering* maps directly onto the course chapters.

The best workshop examples from your own projects are: torbjorn, symphony, C:/dev as an agentic workspace, gchatctl / jiractl as agent-facing CLIs, tlslide for visual verification, reed for markdown artifacts, and D:/workspaces/workshopper itself.

Course chapter mapping to your projects

Chapter 1 — Foundations: agentic loop, LLM vs agent, model vs harness

Strong examples:

  • D:/projects/torbjorn and C:/dev/torbjorn
  • Explicitly describes a fresh-context AI task loop inspired by Geoffrey Huntley's Ralph Loop.
  • Shows the agentic loop in practical form: queue task → spawn agent → observe completion signal → run backpressure → retry or commit.
  • D:/projects/pi_action_ui
  • Tauri UI that scans .elf/*.md action files and runs Pi with --print --no-session @<action-file>.
  • Great “harness vs model” demo: the UI/harness chooses working dir, action files, execution, logs; the model is only one part.
  • D:/projects/pi_auto_runner
  • .pi files as double-click agent scripts. Useful for explaining that agent instructions can become executable workflow artifacts.
  • D:/projects/handsfree-pi
  • AirPods/media-key control of terminal + Wispr Flow. Not an AI model project, but a strong example of lowering human friction around the agentic loop.

Workshop use:

  • Demo torbjorn run or draw its loop as the simplest concrete version of “agents are tools in a loop”.
  • Contrast pi_action_ui / .elf / .pi files with chat-only prompting.

Chapter 2 — Landscape: vibe coding vs agentic engineering

Strong examples:

  • D:/workspaces/workshopper/workshops/vibe-coding-to-agentic-engineering
  • D:/workspaces/workshopper/workshops/vibe-coding-to-agentic-engineering-v2

These already frame the transition from vague prompting to structured agent workflows. The v1 outline includes:

  • LO1: explain what a coding agent is.
  • LO2: turn vague vibe coding into structured agent workflow.
  • LO3: tests/linting/CI/review as backpressure.
  • LO4: decide when to grant or withdraw autonomy.
  • LO5: choose realistic production setup.

Workshop use:

  • Treat this as the canonical workshop scaffold.
  • Use your own projects as examples under each learning outcome.

Chapter 3 — Getting hands dirty: setup, CLI, first loop

Strong examples:

  • D:/projects/pi_auto_runner
  • .pi automation file launcher for Windows.
  • D:/projects/pi_action_ui
  • repeatable Pi actions from a workspace-local .elf folder.
  • D:/projects/handsfree-pi
  • terminal-window control for agent workflows.
  • C:/dev/gchatctl, C:/dev/jiractl, C:/dev/porkctl, C:/dev/brainctl
  • small CLIs with JSON-friendly output, good as first “tools for agents” examples.

Workshop use:

  • Setup exercise: create a tiny workspace with one action file, one AGENTS.md, one CLI command, and one verification command.
  • Show why CLI beats GUI for repeatable agent workflows.

Chapter 4 — Context: AGENTS.md, CLAUDE.md, skills, artifact files, compaction

Strong examples:

  • C:/dev/AGENTS.md and C:/dev/CLAUDE.md
  • Root workspace conventions: _brain, _skills, _env, _docs, _archive, project/tool/skill separation.
  • C:/dev/_brain/README.md
  • Context folder for LLM agents: identity, workflow, tools, goals, voice, projects.
  • C:/dev/_skills/README.md
  • Skill packaging standard: skills are documentation, binaries live on PATH, source and deployed skill copies separated.
  • Project-level instruction files:
  • D:/projects/reed/AGENTS.md
  • D:/projects/slateshow/AGENTS.md
  • D:/projects/tlslide/AGENTS.md
  • D:/projects/symphony/CLAUDE.md
  • D:/projects/siever.ing/AGENTS.md
  • D:/workspaces/workshopper/.pi/skills/workshop-designer/SKILL.md
  • A project-local skill with facilitation gates, checkpoints, and template files.

Workshop use:

  • Show C:/dev as a real “agentic OS” example.
  • Show how _brain, _skills, _env, and project instructions separate context, procedure, secrets, and code.
  • Use workshop-designer as a local skill demo: a skill is not magic; it is a structured how-to file loaded when relevant.

Chapter 5 — Quality and Control: backpressure, tests, subagents, prompt injection, anti-patterns

Strong examples:

  • D:/projects/torbjorn
  • Backpressure commands in .torbjorn/config.toml: build/test/lint gate completion.
  • Agent never decides if work is done; verification does.
  • C:/dev/_gate/README.md
  • Dedicated folder for quality gating, prompt-injection threat models, privilege separation, eval/backpressure, sandboxing, proof-carrying code, property testing.
  • D:/projects/tlslide/AGENTS.md
  • Requires using a project-local Pi extension to start app, screenshot, inspect UI, fetch smoke test, stop server.
  • Very strong example of visual backpressure and “inspect the result yourself”.
  • D:/projects/reed/AGENTS.md
  • Requires bead issue tracking and non-interactive shell flags; good operational guardrail example.

Workshop use:

  • Run a before/after exercise: agent changes UI without screenshot vs with screenshot artifact.
  • Use torbjorn to teach “control flow belongs in scripts/harnesses, not in a giant prompt”.

Chapter 6 — Scaling up: orchestrators, reusable skills, team sharing, background agents

Strong examples:

  • D:/projects/symphony
  • Autonomous coding agent orchestrator connecting Linear issues to Claude Code / Codex.
  • Architecture includes tracker adapter, agent adapters, workflow parsing, workspace lifecycle, logs, dashboard, guardrails.
  • D:/projects/torbjorn
  • Sequential fresh-context task loop with queue, dependencies, reflection, learnings.
  • C:/dev/_skills/README.md
  • Reusable skill distribution pattern with skill-sync, source-of-truth SKILL.md, binary-on-PATH rule.
  • C:/dev/AGENTS.md
  • Defines project/tool/skill layers and workspace conventions.

Workshop use:

  • Compare torbjorn and symphony:
  • torbjorn: simple local queue/backpressure loop.
  • symphony: tracker-driven team/workflow orchestrator.
  • This comparison makes “scaling up” concrete without needing external examples.

Chapter 7 — Harnesses, Models, Providers

Strong examples:

  • D:/projects/symphony
  • Agent adapters for Claude Code / Codex.
  • Workflow and guardrail architecture clearly separates tracker, agent, workspace, guardrail, and logging concerns.
  • D:/projects/torbjorn
  • Model escalation ladder: codex → sonnet → opus; reflection model; retries per tier.
  • C:/dev/_archive/agent
  • Archived custom agent code includes provider resolution for Anthropic / Codex / opencode and model metadata. Useful as a historical/advanced example, not the cleanest workshop demo.

Workshop use:

  • Use torbjorn config to explain provider/model choice as a harness decision.
  • Keep enterprise provider/legal discussion conceptual unless you add a specific Bedrock/Azure/Mistral project example later.

Chapter 8 — Codebase Poisoning

Strong examples:

  • D:/projects/siever.ing/AGENTS.md
  • Static site rules: URL structure, asset organization, archive location, build output.
  • D:/projects/reed/README.md
  • Clean, minimal markdown reader architecture; good “small trusted core” example.
  • D:/projects/ad-in/README.md
  • Data pipeline with canonical outputs, source notes, QA reports, schema roadmap.
  • D:/projects/tlslide/.pi/artifacts/screenshots
  • Many visual artifacts from iterative development. Useful to discuss how a repo teaches the agent through examples, good and bad.

Workshop use:

  • Exercise: give participants a messy mini-repo vs one with a clean AGENTS.md and canonical examples. Compare outputs.
  • Use siever.ing and ad-in to show how explicit structure prevents drift.

Chapter 9 — Autoresearch / closed experiment loops

Direct ML autoresearch examples were not obvious in the scanned projects.

Closest analogues:

  • D:/projects/ad-in
  • Data pipeline with data quality reports and repeatable source-to-schema workflow.
  • D:/projects/tlslide
  • Visual iteration loop with screenshots as empirical artifacts.
  • D:/projects/torbjorn
  • General closed loop: task → agent changes → backpressure → keep/discard/retry.

Workshop use:

  • Present Karpathy autoresearch as the ML-specific version.
  • Then show torbjorn / tlslide as non-ML analogues: autonomous iteration works best with narrow surface area and objective feedback.

Chapter 10 — Vocabulary

Strong examples:

  • D:/workspaces/workshopper/workshops/vibe-coding-to-agentic-engineering-v2/03_outline.md
  • Already has a strong “Vocabulary is leverage” section.
  • C:/dev/AGENTS.md
  • Defines project / tool / skill as separate layers.
  • C:/dev/_skills/README.md
  • Defines source code, deployed skill, binary, env secrets, and sync workflow.

Workshop use:

  • Start workshop with the vocabulary distinction exercise.
  • Ask: “Was this a model failure, context failure, tool failure, harness failure, or eval failure?” This is already in the workshop outline.

Best project examples to use in the workshop

1. D:/projects/torbjorn

Use for:

  • Agentic loop
  • Fresh context
  • Backpressure
  • Model escalation
  • Reflection/learnings
  • Harness vs model
  • Control flow by code instead of prompt

Why it is strong:

  • It is almost a direct embodiment of the course thesis: useful agents emerge from loops, constraints, feedback, and artifacts.

2. D:/projects/symphony

Use for:

  • Scaling up
  • Orchestrator pattern
  • Tracker-driven agent workflows
  • Team/Linear integration
  • Workspace lifecycle
  • Guardrails and metrics

Why it is strong:

  • Shows agentic engineering beyond local coding: tracker → agent → workspace → verification → merge.

3. C:/dev workspace

Use for:

  • Agentic OS
  • Context as architecture
  • Skills
  • Env/secrets separation
  • Project/tool/skill separation
  • Personal/team memory

Why it is strong:

  • This is your own “folder structure as agent architecture” in the wild.

4. C:/dev/gchatctl and C:/dev/jiractl

Use for:

  • Agent-facing CLI design
  • JSON output
  • Non-interactive automation
  • Building tools for agents instead of only humans

Why it is strong:

  • They are simple, practical examples of turning external systems into agent tools.

5. D:/projects/tlslide

Use for:

  • Browser/UI verification
  • Screenshot artifacts
  • Project-local Pi extensions
  • Visual backpressure

Why it is strong:

  • It turns “verify your work” into a concrete habit with artifacts.

6. D:/projects/reed

Use for:

  • Artifact files
  • Markdown-as-interface
  • Minimal tools
  • Agent-readable outputs
  • Operational guardrails via AGENTS.md

Why it is strong:

  • It supports the course’s “plans and artifacts should live in files” message.

7. D:/workspaces/workshopper

Use for:

  • The workshop process itself
  • Project-local skills
  • Folder-based workshop artifacts
  • Learning outcome discipline

Why it is strong:

  • It is the meta-example: you are designing an agentic-engineering workshop using an agentic workspace.

Workshop references already in place

D:/workspaces/workshopper/README.md says the workspace is for designing educational workshops with Pi and can pull source material from D:/course/chapters.

D:/workspaces/workshopper/.pi/skills/workshop-designer/SKILL.md defines a strong workshop design process:

  1. audience profile
  2. schedule chunks / breaks
  3. learning outcomes
  4. supporting ideas
  5. skeleton
  6. next actions

It also correctly prevents premature skeleton generation until outcomes and supporting ideas are reviewed.

Existing workshop folders:

  • D:/workspaces/workshopper/workshops/vibe-coding-to-agentic-engineering
  • D:/workspaces/workshopper/workshops/vibe-coding-to-agentic-engineering-v2

Recommendation: continue from v2. It is more nuanced, especially around vocabulary, context, documentation rot, and workspaces as operating environments.

Recommended workshop storyline

Suggested title:

From Vibe Coding to Agentic Engineering

Suggested arc:

  1. Demystify — An agent is an LLM in a harness with tools and a loop.
  • Example: torbjorn, pi_action_ui.
  1. Name the parts — model, provider, harness, context, tool, eval, agent.
  • Example: workshop v2 vocabulary section.
  1. Shape context — durable files, AGENTS/CLAUDE, skills, project structure.
  • Example: C:/dev, _brain, _skills, workshopper.
  1. Add feedback — tests, lint, screenshots, issue tracking, review gates.
  • Example: tlslide, reed, _gate, torbjorn.
  1. Scale carefully — queues, fresh contexts, orchestrators, tracker integration.
  • Example: torbjorn vs symphony.
  1. Choose stack realistically — harness/model/provider/hosting based on risk and reversibility.
  • Example: torbjorn model ladder; conceptual provider matrix.

Gaps / opportunities

  1. Autoresearch chapter needs a dedicated demo
  • You have analogues, but not a clean ML-style train.py / program.md / metric loop.
  • Consider creating a tiny toy benchmark project for the course.
  1. Provider/compliance examples are mostly conceptual
  • You have model/harness examples, but not much Bedrock/Azure/Mistral/self-hosted implementation.
  • For a DACH/enterprise workshop, create a simple decision matrix handout rather than a code demo.
  1. Prompt injection needs a compact practical demo
  • C:/dev/_gate has theory and references.
  • A workshop would benefit from a small “malicious README / untrusted webpage / tool permission” exercise.
  1. Codebase poisoning could use a before/after repo
  • Your real projects contain examples, but a controlled teaching repo would make the lesson clearer.

Short list of artifacts to prepare next

  1. Pick workshops/vibe-coding-to-agentic-engineering-v2 as the active workshop.
  2. Add a file: 06_examples_from_my_projects.md with the seven examples above.
  3. Create one slide/demo per anchor project:
  • torbjorn: loop + backpressure diagram.
  • C:/dev: agentic OS folder map.
  • gchatctl/jiractl: agent-facing CLI output.
  • tlslide: screenshot verification loop.
  • symphony: team-scale orchestrator diagram.
  1. Create one tiny prompt-injection exercise.
  2. Create one toy autoresearch/benchmark exercise or explicitly mark it as advanced/future.

Bottom line

The course material is already deeply reflected in your projects. The workshop should not feel abstract. Your strongest angle is: “I am not teaching theory I collected; I am teaching the operating model I ended up building.”

The most persuasive examples are the ones where the course principle is visible in files:

  • AGENTS.md / CLAUDE.md
  • .pi/skills/*/SKILL.md
  • .torbjorn/config.toml
  • screenshot artifacts
  • JSON-first CLIs
  • workspace folders as durable context