Building Effective AI Agents Simple, Composable Patterns
summary
Argues that simple, composable patterns beat monolithic frameworks and shows working artifacts: an Audit Agent in production for smart contract auditing, a minimal agent implementation that runs a MinimalAgent loop using explicit tools like searchDocs calculate writeFile, and an evaluator–optimizer loop that iterates generation and critique until quality targets. Demonstrates prompt chaining, routing/classification, parallel code review, and orchestrator-worker flows, with concrete examples (pricing-summary agent that searches, computes 10×12, writes file) and recommends direct API calls for debuggability.
AI-written summary · about
chapters (17)
- 0:00Introduction & Background
- 0:49Observations on Building AI Agents
- 1:26Defining Agentic Systems vs. Workflows
- 2:40When to Use Fully Autonomous Agents
- 4:13Comparing Existing Frameworks
- 5:01Simple Composable Patterns
- 6:26Pattern 1: Prompt Chaining
- 7:21Pattern 2: Routing & Classification
- 9:37Pattern 3: Parallelization
- 11:22Pattern 4: Orchestration & Workflows
- 15:51Pattern 5: Evaluator–Optimizer Loop
- 20:10Minimal Autonomous Agents
- 24:30Example: A Minimal Agent’s Internal Loop
- 29:30Practical Example (Pricing Summary Agent)
- 33:00Combining Patterns & Handling Complexity
- 36:00Kirill’s Current Projects & Wrap-Up
- 37:03Q&A: Multi-Agent Systems vs. Workflows
related talks
27mWhy the Best AI Agents Are Built Without Frameworks (Primitives over Frameworks)
26mBuilding Agentic Workflows: Patterns for Orchestrating Intelligent Systems.
44mBuilding Agentic AI Systems: From Transactions To Conversations
49mAgentic AI Patterns
27mEmerging Patterns Agentic AI
35mAI Agent Control Logic with Pattern Matching
31mAgentic AI architecture patterns
45mAI agents in practice - beyond the prototype