Field notes on building agents that hold up in production.

What I've learned shipping LLM systems for enterprise data teams, explained with diagrams rather than jargon.

1. Sep 8, 2026Context offloading: keep the window small and the work bigLong-running agents drown in their own transcripts. Offloading moves bulky state out to files, stores and sub-agents so the model only sees what matters right now.Context9 min
2. Aug 19, 2026Context is a budget, and every agent is overspendingThe context window is the agent's working memory. How to allocate it, what to compact, and why more tokens usually means worse answers.Context10 min
3. Jul 28, 2026Agent memory: what to remember, where to put it, when to forgetWorking, episodic, semantic and procedural memory for LLM agents, and the write, retrieve and decay policies that keep memory useful instead of noisy.Memory10 min
4. Jul 2, 2026The harness is the productThe model is a function. The harness around it (the loop, tools, permissions, context and recovery) is what makes it an agent. A tour of the parts that matter.Agents11 min
5. Jun 10, 2026MCP from the wire upWhat the Model Context Protocol actually is: hosts, clients, servers, the JSON-RPC handshake, and the design decisions I've learned building MCP clients for enterprise data.Protocols12 min
6. May 14, 2026Tool calling is structured output with consequencesHow an LLM 'calls' a function, why tool design matters more than prompt design, and the patterns that make tool use reliable in production.Agents9 min
7. Apr 22, 2026Putting LLMs inside data pipelines without breaking themLessons from wiring LLMs into Spark and Databricks pipelines: where they belong, where they don't, and how to keep a probabilistic step inside a deterministic system.Data10 min