--- title: "Useful Resources for AI Agents" pubDate: 2026-01-02T00:00:00.000Z tags: - AI - Resources image: /images/blog/useful-resources-for-ai-agents/featured.png --- These resources cover agent frameworks, evaluation, observability, and production engineering. > **Note:** The sources listed here can be used as context when planning an AI agent architecture, for example from NotebookLM or other AI-powered planning tools. These resources provide comprehensive information that can help guide architectural decisions and implementation strategies. ## Anthropic Engineering Blog The [Anthropic Engineering Blog](https://www.anthropic.com/engineering) covers agent architecture, tool use, security, and context engineering. Its posts often include code and case studies from Anthropic's production systems. ## DSPy [DSPy](https://dspy.ai/) replaces brittle prompt strings with structured programs (see my [blog post on DSPy](/post/stop-prompting-start-programming-dspy)). You define the desired input and output as signatures, then use its optimizers to tune the instructions. Modules such as ReAct and ChainOfThought can run across different models. ## MLflow GenAI [MLflow GenAI](https://mlflow.org/genai) tracks and evaluates GenAI applications and agent workflows. Its OpenTelemetry-compatible traces capture prompts, retrievals, and tool calls. It also includes LLM-as-a-judge metrics, a Prompt Registry, and agent versioning that complements Git. MLflow is open source and works across frameworks. ## Measuring Agents in Production The paper ["Measuring Agents in Production"](https://arxiv.org/abs/2512.04123) surveys 306 practitioners and reports 20 case studies across 26 domains. It examines why organizations build agents, how they build and evaluate them, and where development gets difficult. Production agents tend to use simple, controllable approaches: 68% execute at most 10 steps before requiring human intervention, 70% rely on prompting off-the-shelf models instead of weight tuning, and 74% depend primarily on human evaluation. ## The 2025 AI Engineer Reading List [The 2025 AI Engineer Reading List](https://www.latent.space/p/2025-papers) from Latent.Space collects approximately 50 papers across 10 fields in AI engineering: LLMs, Benchmarks, Prompting, RAG, Agents, CodeGen, Vision, Voice, Diffusion, and Finetuning. Each entry explains why the paper matters, with an emphasis on practical applications. ## Backpressure Is All You Need [Backpressure Is All You Need](https://www.lucasfcosta.com/blog/backpressure-is-all-you-need) by Lucas F. Costa explains how automated feedback loops can make longer, unattended coding-agent sessions safer and more useful. It treats tests, types, linters, benchmarks, review agents, manual verification, and pull-request monitoring as forms of backpressure that force an agent to catch and correct problems before involving a human. The article offers a practical blueprint for moving human review toward higher-level architecture and design decisions instead of repetitive correctness checks. [Visual Attention Variants](https://magazine.sebastianraschka.com/p/visual-attention-variants) by Sebastian Raschka.