The Agent Stack Behind Reliable Agentic Systems

Agents fail in production for reasons that rarely show up in a demo: lost context, duplicated work, runs that die halfway and start over. The model isn't the problem. The harness is. The agent harness is the engineering layer wrapped around the model: it plans, fans work out, verifies results, checkpoints progress, and decides what enters the context window. Reliability lives in this layer, and every one of those capabilities resolves to a memory operation. This session is a working tour of the harness's memory surfaces, the places where memory engineering decides whether an agentic application can be trusted with real work.

What we'll cover:

• The state of agentic applications in 2026: autonomous systems that don't just run automations, they build them, and the arrival of first-party harnesses like dynamic workflows in Claude Code

• Surface one, injection: getting memory into a live agent with Anthropic's mid-conversation system messages, delivering operator-level priority with zero cache invalidation

• Surface two, coordination: rebuilding the dynamic-workflows pattern as a lightweight custom harness where task claims, findings, and checkpoints live in shared agent memory

• Surface three, persistence and recall: shared state and vector search in Oracle AI Database, keeping parallel agents coherent and interrupted runs resumable

The stack we build on: PALO (Python, Anthropic, LangChain, and Oracle)

You'll leave with a reference architecture, the code to run it, and a checklist of the memory surfaces your own harness needs to cover before you call it reliable.

Duration: 90 minutes