agentic-ai-engineering/course/SLIDES.md

161 lines
8.9 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# Slide Deck Outlines
Presentation outlines for all 8 modules. Each module has ~10-15 slides.
---
## Module 1: Foundations (~12 slides)
| Slide | Title | Content |
|-------|-------|---------|
| 1 | Title | Agentic Engineering: From Foundations to Production |
| 2 | What Makes an Agent? | LLM + Tools + Loop = Agent. Diagram of the 3 components. |
| 3 | The Harness vs The Model | Table comparing what matters. "The harness IS the product." |
| 4 | 5 Subsystems of a Harness | Instructions, Tools, Environment, State, Verification. Diagram. |
| 5 | 4 Dimensions of Control | Context, Model, Prompt, Tools. Which gives you leverage? |
| 6 | The Repository IS the Spec | Single source of truth. CLAUDE.md, init.sh, feature_list.json. |
| 7 | Decision Framework | 4-question filter. When to use agents, when not to. |
| 8 | The Agent Loop | Think → Act → Observe → Repeat. Pseudocode. |
| 9 | Tool Calling Deep Dive | Anatomy of a tool. The reasoning parameter. Forced vs auto. |
| 10 | Vibe Coding vs Agentic Engineering | 5 hard rules. Risk compounds with runtime. |
| 11 | Do You Trust Your Agents? | The course thesis. Discussion question. |
| 12 | Module Summary | Key takeaways. Prep for Lab 1.8. |
---
## Module 2: Architecture (~15 slides)
| Slide | Title | Content |
|-------|-------|---------|
| 1 | Title | Agent Architecture |
| 2 | The Four Pillars | Tools, Loop, Context, Memory. Diagram. |
| 3 | Tool Design Patterns | MCP vs CLI vs Script vs Skills. Context cost matrix. |
| 4 | LSP MCP Server | Symbol-level search. Diagram. When grep is too slow. |
| 5 | Agent Loop Variants | 5 levels: Simple → Tool-Use → Task-Completion → Sub-Agent → Multi-Agent |
| 6 | Skills System Deep Dive | Path-scoped skills, subdirectory CLAUDE.md, plugin distribution. |
| 7 | Agent-Readable Workspace | init.sh, feature_list.json, clean state. The discovery pattern. |
| 8 | Agent Loop Variants Detail | Code examples for each level. |
| 9 | Context Window Management | Sliding window, summarization, structured outputs. Hybrid approach. |
| 10 | Memory Patterns | Mental models, expertise files, scratch pads, session logs. |
| 11 | The Reasoning Parameter | Force chain-of-thought on every tool call. Code example. |
| 12 | Codebase Architectures | Atomic, Layered, Pipeline, Vertical Slice. Comparison. |
| 13 | Architecture Decision Tree | Which pattern for which project? |
| 14 | Module Summary | Key takeaways. Prep for labs. |
---
## Module 3: Safety & Security (~14 slides)
| Slide | Title | Content |
|-------|-------|---------|
| 1 | Title | Safety & Security |
| 2 | Prompt Injection (L0) | ACIP defense. Attack vectors: direct, indirect, role-play. |
| 3 | Why Bash Is the Problem | Single funnel for every destructive action. Diagram. |
| 4 | The Probability Math | 1% × 100 turns = 63%. The chart that changes everything. |
| 5 | The 6-Level Security Ladder | L0 (ACIP) → L1 (skill) → L2 (system prompt) → L3 (blacklist) → L4 (whitelist) → L5 (no bash) |
| 6 | L3 Marque Break | Agent writes cleanup.py → runs python cleanup.py. The demo. |
| 7 | Damage Control | 3 access levels: zero-access, read-only, no-delete. YAML config. |
| 8 | Hook Architecture | 13 lifecycle events. PreToolUse, PostToolUse, Stop, etc. |
| 9 | The Verifier Pattern | Builder + Verifier over Unix socket. Diagram. |
| 10 | Confidence Ladder | PERFECT → VERIFIED → PARTIAL → FEEDBACK → FAILED |
| 11 | Defense-in-Depth Stacking | All 6 levels stacked. No single point of failure. |
| 12 | Security Decision Tree | Which level for which deployment? |
| 13 | Module Summary | Key takeaways. Prep for labs. |
---
## Module 4: Multi-Agent Orchestration (~18 slides)
| Slide | Title | Content |
|-------|-------|---------|
| 1 | Title | Multi-Agent Orchestration |
| 2 | Why One Agent Is Not Enough | Context ceiling, capability ceiling, reliability ceiling. |
| 3 | 3 Orchestration Patterns | Dispatcher, Pipeline, P2P. Comparison diagram. |
| 4 | P-Threads | N agents in parallel. Pick best result. mprocs example. |
| 5 | F-Threads | N agents → Judge → Winner. Your UI Agents system. |
| 6 | Depth-2 Delegation | Orchestrator → Leads → Workers. Never execute. |
| 7 | Agent Experts That Remember | Mental models. Self-improve commands. Compounding knowledge. |
| 8 | Domain Locking | Per-agent permissions. Read/write/delete scope. |
| 9 | TillDone Task Discipline | Task list gating. Live progress. Nudge on incomplete. |
| 10 | Agent Chains | YAML-defined pipelines. $INPUT/$ORIGINAL variables. |
| 11 | P2P Communication | Pi-to-Pi. 4 tools. Hop limits. Self-healing. |
| 12 | MCP Agent Mail | Jeff's alternative: inboxes, threads, file leases. |
| 13 | Service Connectors | Twitter, GitHub, Linear, Slack via MCP. One per service. |
| 14 | Conversation Awareness | Shared JSONL logs. Cross-agent context. |
| 15 | CEO Board System | 8 agents. Adversarial debate. Fact verification. |
| 16 | UI Agents System | 12 agents. Brand → Product → Tree → Branch → Leaf. |
| 17 | Orchestration Decision Tree | Which pattern for which problem? |
| 18 | Module Summary | Key takeaways. Prep for labs. |
---
## Module 5: Production Patterns (~14 slides)
| Slide | Title | Content |
|-------|-------|---------|
| 1 | Title | Production Patterns |
| 2 | Production Is Different | Deterministic vs non-deterministic. Fixed vs variable cost. |
| 3 | CI/CD for Agents | Golden datasets. Regression testing. Prompt versioning. |
| 4 | The 5-Tool Production Stack | Your architecture: Claude Code, Pi, OpenCode, Hermes, OpenClaw. |
| 5 | Stack Diagram | agent-mux → mprocs → psmux → dmux. Full stack. |
| 6 | The Agent Manager Role | 5 areas of ownership. 90-day playbook. |
| 7 | Shadow Deployments | Run new agent alongside prod. Compare before switching. |
| 8 | Rollback Strategies | Prompt + model + parameters + config. Version-locked configs. |
| 9 | Observability | SQLite tracing. Every tool call logged. Decision tracing. |
| 10 | Cross-Provider Session Search | Jeff's tool. Index sessions across 11 providers. |
| 11 | Alerting | Agent-specific signals: tool loops, cost spikes, grinding. P0/P1/P2 routing. |
| 12 | Deployment Modes | Local trusted, authenticated private, authenticated public. |
| 13 | Cost Control | Budgets per agent/session/task. Warnings vs hard stops. |
| 14 | Module Summary | Key takeaways. Prep for labs. |
---
## Module 6: Economics & Evaluation (~12 slides)
| Slide | Title | Content |
|-------|-------|---------|
| 1 | Title | Economics & Evaluation |
| 2 | The Compute Advantage Equation | (Compute Scaling × Autonomy) ÷ (Time + Effort + Cost). |
| 3 | Your Stack's Compute Advantage | Real numbers for each tool in your stack. |
| 4 | LLM Pricing Landscape | 100x range. Gemini Flash ($0.15/M) to Claude Opus ($15/$75/M). |
| 5 | Cascade Routing | Cheap model for simple steps, expensive for complex. 66% savings. |
| 6 | Cost Per Session Math | The 3x rule. Retries, tool overhead, context growth. |
| 7 | Jeff Emanuel Case Study | 52 subs, $12K/mo, 85K commits. What's his CAC? |
| 8 | Agent Evaluation Metrics | pass@k, tool call accuracy, cost per task, loop efficiency. |
| 9 | Automated Evaluation | Golden datasets. VCR-style recording. Regression detection. |
| 10 | A/B Testing Agents | Canary deployments. 5% traffic split. Compare success rate, cost. |
| 11 | Human Evaluation | Spot-check 5-10%. Catches 2-3x more failures than automated. |
| 12 | Module Summary | Key takeaways. Prep for labs. |
---
## Module 7: Advanced Topics (~10 slides)
| Slide | Title | Content |
|-------|-------|---------|
| 1 | Title | Advanced Topics |
| 2 | Autoresearch | Run → Measure → Log → Decide. Self-improving agents. |
| 3 | Integrity Guards | Code hashing (anti-grinding). Median over best. Test set integrity. |
| 4 | Meta-Agents | Agents that build agents. Parallel research experts. |
| 5 | Beyond MCP | MCP vs CLI vs Script vs Skills. Context cost matrix. |
| 6 | Mac Mini Agent | Physical sandbox. Steer + Drive + Listen + Direct. |
| 7 | Always-On Agents | Voice→CLI bridge. Job server pattern. Heartbeat execution. |
| 8 | Framework Comparison | Paperclip vs Multica vs Cabinet. When to use which. |
| 9 | The Future | Agentic Coding 2.0. Agents conducting agents. |
| 10 | Module Summary | Key takeaways. Prep for labs. |
---
## Module 8: Capstone (~8 slides)
| Slide | Title | Content |
|-------|-------|---------|
| 1 | Title | Capstone: Production Multi-Agent System |
| 2 | Project Options | Brand Monitor, Code Review Pipeline, Strategic Decision Board. |
| 3 | Phase 1: System Design | Architecture diagram. Agent roster. Data model. Cost estimate. |
| 4 | Phase 2-3: Build + Integrate | Agent implementation. Mental models. Wiring agents together. |
| 5 | Phase 4-5: Security + Test | L3 minimum. Golden dataset. pass@k evaluation. |
| 6 | Phase 6-7: Deploy + Review | CI/CD. Shadow deploy. Red-team. Retrospective. |
| 7 | Pass Criteria | What constitutes a passing project. Rubric. |
| 8 | Next Steps | After the capstone. Continuing education. Community. |