Agent Engineering: routing, prompts, production, and domain stacks

New Alibaba AI framework skips loading every tool, cutting agent token use 99%. Alibaba’s SkillWeaver builds execution graphs and performs iterative skill retrieval to avoid loading every tool, cutting token use by ~99% while improving multi-tool routing accuracy. For outcome engineers this is a clear pattern: lazy skill retrieval and execution graphs let you scale multi-tool agents without token- or latency-explosion (Principles 09, 11, 06).

Three Ways LLM Pipelines Fail in Production That Staging Will Not Catch. The piece identifies schema drift, retrieval-context poisoning, and hidden embedding lifecycle bugs as causes of silent production failure and prescribes prompt/schema versioning, chunk hashing, and atomic embedding updates. This matters because reproducible, auditable retrieval + schema practices are foundational to stable outcome pipelines in production (Principles 06, 11, 14).

Using DSPy to evaluate and improve Datasette Agent’s SQL system prompts. Simon Willison uses DSPy to find prompt weaknesses and guide schema-listing changes that improve Datasette Agent’s SQL reliability. Outcome engineers should adopt programmatic prompt-evaluation and schema-driven grounding to reduce hallucinations and make agent behavior auditable (Principles 06, 16).

Vercel’s Andrew Qu on why agents are a new kind of software. Vercel describes eve as a developer-first agent platform built around prescriptive primitives — context, tools, resumability, and subagents — to make agents testable and composable. That framing matters for teams designing agent platforms: bake resumability and explicit context primitives into your runtime so agents behave like maintainable software (Principles 09, 06).

Trunk Tools’ stack cut document review from 60 days to 10 by ditching general-purpose models. Trunk Tools replaces general LLMs with a perception–semantics–agents stack and an agent-ready knowledge graph to dramatically speed domain workflows. This is a concrete playbook for outcome engineering: align perception layers, specialized semantic indexes, and orchestrated agents to deliver measurable outcome improvements (Principles 06, 11, 09).