Skip to Content
View all events

Advanced Harness Engineering

Published by O'Reilly Media, Inc.

Advanced content levelAdvanced

Design and build agentic workflow and deep research harnesses

What you’ll learn and how you can apply it

  • Design a durable workflow harness that combines checkpointing, idempotent task execution, approval gates, and auditable execution logs
  • Build a workflow agent in LangGraph that persists state, resumes from checkpoints after a simulated failure, and uses procedural and semantic memory for SOPs and business rules
  • Design a deep research harness with an orchestrator-worker architecture, a shared evidence pool, and a memory consolidation pipeline
  • Build a parallel multi-agent deep research system that accumulates knowledge across sessions and demonstrates measurable improvement session over session

Course description

Workflow agents and deep research agents are two of the hardest application modes to get right in production. Workflow agents need to run for minutes or hours, survive failures, persist state across interruptions, and produce audit trails that hold up under review. Deep research agents need to coordinate parallel subagents, share evidence across a growing knowledge base, consolidate findings into durable knowledge, and resolve contradictions before they compound. The harness requirements for each are distinct, and the architectural choices you make early tend to determine whether the system can scale or quietly collapse as tasks get longer.

This two-hour course with Richmond Alake applies a memory-first, component-driven approach to the workflow and deep research application modes. You’ll design a durable workflow harness using LangGraph-style checkpointing, encode SOPs and business rules into the appropriate memory types, and build approval gates that keep a human in the loop without breaking determinism. You’ll then design a parallel, multi-agent deep research harness with an orchestrator and worker agents, a shared evidence pool, and a memory-consolidation pipeline that turns episodic traces into durable semantic knowledge.

This live event is for you because...

  • You’re a software engineer, ML engineer, or AI engineer who has built agent prototypes and is taking on longer-running or higher-stakes workloads.
  • You’re a technical lead or architect designing workflow automation or research tooling that needs to run reliably over hours or days.
  • You work with LangGraph, CrewAI, LlamaIndex, the Claude Agent SDK, or similar frameworks and want to apply harness engineering principles to multi-agent systems.
  • You want to move beyond single-shot agents and build systems that run reliably over long horizons.

Prerequisites

  • A Python 3.11+ environment with a recent agent framework installed (LangChain, LangGraph, or the Claude Agent SDK)
  • An API key for at least one frontier model provider (Anthropic, OpenAI, or equivalent)
  • A local database available for memory persistence (Oracle AI Database)
  • A code editor with notebook support, and Git installed for cloning the course repository (shared before the session)
  • Working knowledge of Python, including async patterns and basic API development
  • Familiarity with LLM APIs and prompt-level agent behavior (tool calling, system prompts, structured outputs)
  • Familiarity with at least one agent framework (LangGraph preferred)
  • Comfort with graph-based state management (helpful but not required)
  • Exposure to vector search and semantic retrieval concepts

Recommended follow-up:

Schedule

The time frames are only estimates and may vary according to how the class is progressing.

Building the workflow mode agent harness (50 minutes)

  • Presentation: Why workflow mode needs durable execution; how checkpointing differs from runtime-managed durability; when sequential, orchestrator-worker, or parallel architectures fit; encoding SOPs and business rules into procedural and semantic memory; treating step outputs as checkpoint writes that enable resumability; sandbox design for workflows, covering idempotent tasks, deterministic replay, tool scope boundaries, and approval gates with auditable logs built on episodic memory
  • Hands-on exercise: Build a compliance reporting workflow with LangGraph checkpointing, a human approval gate, and the four memory types wired in; trigger a mid-pipeline failure and resume from the last persisted checkpoint
  • Break

Building the deep research mode agent harness (50 minutes)

  • Presentation: Why deep research needs a parallel multi-agent architecture; when multi-agent systems help and when they amplify errors; designing the shared evidence pool as a memory substrate, covering provenance, deduplication, and contradiction detection; memory consolidation pipelines that turn episodic traces into durable semantic knowledge; context engineering across a growing knowledge store, recursive subtask decomposition, semantic caching for search and synthesis, and human-in-the-loop steering for research direction
  • Hands-on exercise: Build a competitive intelligence harness with an orchestrator agent, parallel worker agents, a shared evidence pool, and a consolidation pipeline; run two sessions back to back and measure how first-session findings improve the second

Wrap-up and Q&A (10 minutes)

Your Instructor

  • Richmond Alake

    Richmond Alake is director of AI developer experience at Oracle, where he leads developer relations for the Oracle AI Database and pioneers the discipline of memory engineering, a term he coined to describe the structured design of agent memory systems. He runs the #100DaysOfAgentMemory LinkedIn series, which has reached tens of thousands of AI practitioners.

    Richmond has published over 200 technical articles with more than one million views across Medium, NVIDIA’s technical blog, and other industry publications, covering topics from agent memory systems to production LLM application development. He has spoken at the AI Engineer World’s Fair, DeveloperWeek, Confluence, and Monki Gras. Previously, he served as an AI/ML developer advocate at MongoDB, where he built MemoRizz, an open source Python agent memory library, and codeveloped a DeepLearning.AI course on prompt compression and query optimization. Richmond holds a master’s degree in computer vision, machine learning, and robotics from the University of Surrey.

    linkedinXlinksearch

Skills covered

  • AI Agents
  • Engineering