No-Code Agents with OpenAI AgentKit

Explore OpenAI's announcements including Agent Kit and App SDK, signaling a shift to an agent-driven web. Understand Agent Kit components: Agent Builder for no-code agent design, ChatKit for conversational AI embedding, Evals for performance testing. Discuss the vision of a post-app web where intelligent agents handle user interactions, integrating services via App SDK. Examine implications for users, developers, businesses, and the web structure. Review competition from Google, Microsoft, Meta, Anthropic, and startups in building AI agent infrastructure.

Introduction to agentic organizations where humans and AI agents collaborate closely, representing a paradigm shift akin to industrial to digital transitions. Address the gen AI paradox: why current AI efforts yield limited gains and the need for vertical integration into workflows. Define agentic organizations with AI agents planning, coordinating, and acting autonomously, while humans focus on oversight and strategy. Discuss value delivery through faster execution, real-time adaptation, personalization, and resilience. Provide guidance on strategic focus, workflow redesign, building capacity, leadership involvement, and pilot projects. Highlight risks like autonomous sprawl, governance needs, technical debt, cultural resistance, and implementation challenges.

The AgentKit stack at a glance: Agent Builder (visual), Connector Registry (admin), ChatKit (UI), Evals (quality), Guardrails (safety). Concepts in human terms: tasks, tools, multi-step flows, agent handoffs, memory/sessions. Demo tour of Agent Builder: canvases, nodes, versions, templates; publish/preview lifecycle. Lab: Clone a template, customize instructions, add an approval step, run test conversations.

What to put in vs. link to; file search basics; connector options (e.g., Drive, SharePoint, Teams) and MCP servers. Admin view of Connector Registry; roles & governance; enabling connectors for a workspace. Data handling patterns: least-privilege, redaction, auditability. Lab: Attach a small policy pack (PDFs/Docs), configure retrieval, and test relevance safely.

Drag-and-drop nodes: tools, file search, guardrails, decision/branching, human-in-the-loop. Multi-agent patterns via handoffs—when and why to split responsibilities. Versioning & change logs; rollback and safe launches. Lab: Build a 5–7 node workflow from a blank canvas; add a human approval and a fallback path.

Shipping a usable interface without front-end work: embed options and theme/brand tweaks. Sharing with pilot users; capturing transcripts and feedback for iteration. Lab: Deploy your agent’s chat UI, set a custom name/avatar, and invite 3 pilot testers.

Designing a simple eval dataset from real tickets/prompts. Trace grading: grading whole runs to spot brittle steps. Prompt optimization: generate improved prompts from grader + human annotations. Third-party model comparisons (what, why, when). Lab: Create a 20-case eval, run it, and apply one optimization round. Re-run and compare.

Reading traces; identifying tool-call loops and dead ends. Lightweight A/Bs: instruction tweaks and guardrail thresholds. When RFT helps (conceptual only) and how to scope an RFT request with your tech team. Lab: Use traces to remove one redundant step and reduce tokens/time per task.

The Guardrails Wizard: select checks (moderation, jailbreak, PII, hallucinations) and set policies—no code. Human approvals & audit trails; rollout controls and change management. Lab: Add PII masking + jailbreak detection to your agent, document your policy, and re-test.

Finalize workflow, run evals, harden guardrails, and polish the ChatKit UI. Write a one-page “launch note”: scope, metrics, SLAs, rollback plan. Capstone demo: 5-minute live run + Q&A; submit your Launch Note and a 14-day iteration plan.