# HARTRO Blog

> Practical guides and research on choosing, building, and evaluating useful AI.

Canonical URL: https://hartro.com/blog/

- [Is an AI project worth the cost? A worked business case](https://hartro.com/blog/ai-project-cost-and-roi.md) (2026-10-21): Calculate AI project ROI with a worked example covering adoption, human review, integration, running costs and the difference between capacity and cash savings.
- [How to choose your first AI workflow](https://hartro.com/blog/choose-first-ai-workflow.md) (2026-10-21): Compare AI workflow ideas, identify blockers and choose a small first test with a clear owner, useful baseline and manageable consequences of error.
- [How to decide whether an AI pilot worked](https://hartro.com/blog/evaluate-ai-pilot.md) (2026-10-21): Set AI pilot success criteria around real work: outcome quality, human effort, adoption and permitted actions. Use the evidence to continue, change or stop.
- [What AI hosting companies actually sell](https://hartro.com/blog/what-ai-hosting-companies-sell.md) (2026-10-21): Open models make weights available. Serving them economically depends on traffic, cache reuse, and the capacity to stay fast under load.
- [Reasoning with less text](https://hartro.com/blog/hybrid-reasoning-agents.md) (2026-10-20): HybridAgent combines latent computation with text-based tool use. The useful questions are what it saves, what remains observable, and what it takes to run.
- [What does a forward deployed engineer actually do?](https://hartro.com/blog/fdes-vs-consultants.md) (2026-10-19): What does a forward deployed engineer do, and how does the role differ from consulting? A look at AI delivery, product development, and their changing economics.
- [When an agent says it's done](https://hartro.com/blog/agent-completion-evidence.md) (2026-10-18): A coding agent's completion report should link its claims to observable results. Verification loops make failures visible before the work reaches the reviewer.
- [What does it mean for AI to learn from your business?](https://hartro.com/blog/ai-learning-from-business.md) (2026-10-17): Using AI for six months does not guarantee it improves. The useful question is what changes: its source knowledge, its working procedures, or the model itself.
- [The future of skills is learning from work](https://hartro.com/blog/the-future-of-skills.md) (2026-10-15): Enterprise AI should retain useful lessons from work, turn them into tested procedures, and keep improving them. WikiSkill and OpenWiki show parts of that future and the work still required.
- [What an agent skill actually is](https://hartro.com/blog/what-agent-skills-are.md) (2026-10-13): What goes in SKILL.md, how an agent finds it, and where prompts, tools, and memory fit.
- [What remains when AI can write the code](https://hartro.com/blog/coding-and-software-engineering.md) (2026-10-10): The strongest coding agents change the economics of implementation. They leave difficult questions about requirements, architecture, evidence, and ownership.
- [Give the agent a definition of done](https://hartro.com/blog/evaluation-driven-development.md) (2026-10-08): Evaluation-driven development turns an ambiguous AI project into a sequence of testable improvements. The hard work is deciding what the tests should mean.
- [Autoresearch is moving into the systems that run AI](https://hartro.com/blog/autoresearch-next-frontier.md) (2026-10-06): Agents can optimize GPU kernels and inference systems. I care about whether those changes make the full workload faster, and where an LLM improves on conventional search.
- [What autoresearch taught us about skin analysis](https://hartro.com/blog/autoresearch-skin-analysis.md) (2026-10-04): An AI agent explored wrinkle and pore models, trained small networks, and built review tools. The difficult part was deciding what our labels could tell us.
- [The agent harness is becoming a runtime](https://hartro.com/blog/agent-harness-runtime.md) (2026-10-02): OpenCode 2.0, Pi 1.0, Codex, and Claude Code point toward a different question: how should useful work survive beyond a single conversation?