Advanced Agentic Coding - Claude Code

Objectives

• Explain the fundamentals of AI Engineering: tokens, embeddings, RNNs, Transformers
• Dissect the anatomy of a coding agent: context, tool calling, the ReAct loop
• Build a custom tool and an autonomous agent with the Claude Agent SDK

What's covered

• AI Engineering fundamentals: tokens and tokenization, embeddings, RNNs, Transformers and attention
• How an agent builds and manages its execution context (filesystem exploration)
• Tool calling: from text to a structured action request
• The ReAct agentic loop: Reasoning → Act (tool call) → Observe
• Common components: model, runtime, tools, memory, observation quality
• Building with the Claude Agent SDK: primitives (messages, hooks, tools)
• Defining a custom tool: input schema, handler, return value
• Wiring the tool and driving the loop via hooks (PostToolUse), exposing it as an SDK or MCP
• Exercise: build a custom agent with the TypeScript (or Python) SDK and a bespoke tool

Objectives

• Distinguish orchestration patterns: sequential, parallel, deterministic (script) vs model-driven
• Set up loop engineering with /loop, /goal and /schedule
• Break a feature down into subtasks orchestrated through dynamic workflows

What's covered

• Recap on subagents: context isolation, specialization, parallelism
• Orchestration patterns: sequential, parallel, deterministic (script) vs model-driven
• Decision table: regular turn / subagent / workflow
• Loop engineering with /loop: time-based polling, CI, deployment, review, exit condition
• Loop engineering with /goal: a measurable finish line proven by evaluation
• Limits: scope confined to the session, durability via /schedule
• Dynamic workflows: scripted orchestration, fan-out, branching, deterministic aggregation
• Primitives: agent, parallel, pipeline, phase; structured output by schema
• Exercise: break a feature down into subtasks and orchestrate it across multiple agents

Objectives

• Define a skill as a contract: name, description, Markdown body, files
• Identify why a skill fails: ambiguity, cost, latency, drift
• Measure trigger accuracy against task performance

What's covered

• A skill as a contract: name, description, Markdown body, files
• The description as a triggering surface (too vague, too broad or too narrow)
• Why skills fail: ambiguity, cost, latency, drift
• Trigger accuracy vs task performance
• Test harness (workflows, subagents, worktrees, JSONL) and key metrics
• Exercise: evaluate a translation-checking skill (ground truth from locale files, WITH/WITHOUT runs, JSONL parsing)
• Measure tool calls, tokens, false positives/negatives, diff violations, trigger accuracy
• Exercise: improve a skill within bounds using /goal until a convincing npm run smoke
• Constrain the goal: 0 false positives/negatives, 0 violations, file scope, stop at 6 turns

Objectives

• Optimize token consumption across three levers: retrieval, context/input, output
• Set up a local knowledge base with retrieval (QMD/MCP, BM25 + semantic)
• Assemble the day's skills into a complete end-to-end workflow

What's covered

• Three optimization levers: retrieval, context/input, output
• Retrieval: local knowledge base with QMD (MCP), BM25 + semantic, Karpathy's Wiki pattern
• Context/input: compressing command output with RTK, principles and trade-offs
• Output: reducing output tokens with Caveman, principles and trade-offs
• Exercise: assemble the day's skills into an end-to-end workflow
• Set up a knowledge base with a provided project's documentation
• Integrate retrieval into the orchestration workflow built earlier
• Wrap-up: recap of key concepts, Q&A and experience sharing