AI 2026-10-08 5 videos

Building Reliable Context for Production AI Agents

Reliable agents need structured context, deterministic execution, visible uncertainty and human approval where mistakes carry real consequences.

Summary

Agent reliability depends on architecture around the model: Reducto limits tools and exposes uncertainty, while production harnesses separate LLM planning from deterministic execution with approvals and recorded side effects. Neo4j turns memory into a typed context graph with entity resolution, decision traces and executable skills, while Postman applies a similar graph approach to APIs, grounding service relationships in code and telemetry. Reducto reports that its sales team adopted the rebuilt system, Postman found API-focused evaluations used about half or a third as many tokens as pure code search, and Neo4j found that typed execution graphs improved task success on SkillsBench. The supplied record for “Your AI Agent Has No Nervous System” contains no summary or key points, so it adds no verifiable detail.

The videos

Why We Deleted Our MCP Server and Rebuilt It — Abhi Arya, Reducto ↗ AI Engineer

Reducto rebuilt its MCP server around fewer, work-shaped tools, snapshots, confidence scores, bounding boxes, evals and an ask-user tool after its endpoint-per-tool design let agents create opaque workflows, with the company having processed over 3 billion documents in 3 years.

Your AI Agent Has No Nervous System — Matt Gibiec, Dynatrace ↗ AI Engineer

The provided record gives no summary or key points for “Your AI Agent Has No Nervous System,” leaving its subject matter unspecified.

Turning Agent Memory Into Skills That Work — Will Lyon, Neo4j ↗ AI Engineer

Neo4j turns agent memory into a typed context graph with entity resolution and decision traces, then uses its APE protocol to make skills executable graphs; applying the protocol to SkillsBench skills significantly increased task success.

We Mapped 115 Microservices for Our Coding Agents — Kamalakannan Nandagopal, Postman ↗ AI Engineer

Postman mapped 115 microservices into an API context graph grounded in code and production telemetry, while evaluations for discovery, redesign and impact assessment used about half or a third as many tokens as pure code search.

Brains vs Hands: How to Run AI Agents Safely in Production — Viren Baraiya ↗ AI Engineer

Production agent harnesses let the LLM plan while deterministic code executes durable, idempotent steps with approval gates, treating runtime-assembled workflows as late-bound sagas.