<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Blog: Notes on production AI systems | Soba Labs</title><description>Essays from Soba Labs on agentic workflows, operating state, knowledge systems, and the craft of getting AI into real business processes.</description><link>https://sobalabs.ai/</link><language>en-us</language><item><title>Context engineering in deep agents: the four levers</title><link>https://sobalabs.ai/blog/context-engineering-in-deep-agents/</link><guid isPermaLink="true">https://sobalabs.ai/blog/context-engineering-in-deep-agents/</guid><description>A deep agent&apos;s ceiling is its context window. The harness ships four separate levers, and knowing which one is working decides whether a long run finishes.</description><pubDate>Mon, 27 Jul 2026 00:00:00 GMT</pubDate></item><item><title>Deep agents in production: what the harness gives you, and what you still own</title><link>https://sobalabs.ai/blog/deep-agents-in-production/</link><guid isPermaLink="true">https://sobalabs.ai/blog/deep-agents-in-production/</guid><description>LangChain&apos;s deepagents harness ships planning, subagents, and a filesystem out of the box. Production still asks you to own the human gate, the evals, and the cost.</description><pubDate>Mon, 20 Jul 2026 00:00:00 GMT</pubDate></item><item><title>How we halved Claude Code token usage with RTK AI, an open source CLI proxy</title><link>https://sobalabs.ai/blog/halving-claude-code-token-usage-with-rtk/</link><guid isPermaLink="true">https://sobalabs.ai/blog/halving-claude-code-token-usage-with-rtk/</guid><description>We routed our coding agents&apos; shell commands through RTK AI, an open source CLI proxy. Verified result: 49.6% of tokens saved across 1,457 commands.</description><pubDate>Mon, 13 Jul 2026 00:00:00 GMT</pubDate></item><item><title>Stop asking the LLM how confident it is</title><link>https://sobalabs.ai/blog/deterministic-confidence-gates/</link><guid isPermaLink="true">https://sobalabs.ai/blog/deterministic-confidence-gates/</guid><description>Self-reported LLM confidence is miscalibrated and clusters near certainty. Gate on a signal you can measure, then escalate the exceptions to a person.</description><pubDate>Mon, 06 Jul 2026 00:00:00 GMT</pubDate></item><item><title>Coding is no longer the bottleneck</title><link>https://sobalabs.ai/blog/coding-is-no-longer-the-bottleneck/</link><guid isPermaLink="true">https://sobalabs.ai/blog/coding-is-no-longer-the-bottleneck/</guid><description>AI made writing code cheap. Architecture, verification, security, and quality are the new constraints, and that is where the real work moved.</description><pubDate>Thu, 25 Jun 2026 00:00:00 GMT</pubDate></item></channel></rss>