Agentic systems need infrastructure around the model as they move from single prompts to tool calling, memory, retrieved context, and longer workflows.

User state has to stay consistent across replicas, tool usage has to stay within limits, repeated work has to be avoided, and latency has to stay low under load.

In this session, I’ll walk through harness engineering for agentic systems and the core patterns that make agents easier to run in production. We’ll cover:

• How agents evolved from prompting to tool calling, context engineering, and harness engineering
• How session management keeps user state consistent across requests
• How rate limiting protects APIs, LLM calls, tools, tenants, and downstream services
• How fixed windows, sliding windows, and token buckets handle different traffic patterns
• How caching reduces latency, repeated provider calls, infrastructure load, and cost
• How cache aside, write through, write behind, and prefetching fit different workloads

We’ll use a stock analysis agent as the running example and show where these patterns matter when agents call external APIs, serve multiple users, and run across multiple replicas.

Redis will appear as one practical implementation layer for shared state, rate limits, and caching. The main focus is the operating model: how to give agentic systems predictable state, controlled usage, faster responses, and fewer repeated calls in production.

Participants will leave with a practical mental model for harness engineering and the first reliability patterns every production agent needs: session management, rate limiting, and caching.

This session covers harness engineering patterns for agentic systems: session management, rate limiting, and caching. We’ll use a stock analysis agent to show how agents keep user state consistent, control tool and API usage, reduce repeated work, lower latency, and run predictably across replicas, with Redis as the implementation layer.