Field notes

Field notes on building,shipping, and funding AI.

Practical writing from the XY Space team. The Academy is our hands-on tutorials for non-developers building production-grade systems, and AI Space is our technical wing: open-model comparisons and setup guides for teams taking back AI and data sovereignty.

All posts

01Funding

IRAP Funding for Canadian Corporations: Get Up to 80% of Your R&D Salaries Back

How the NRC's IRAP program funds small and mid-size R&D projects for Canadian businesses, what makes a project fundable, and how it stacks with SR&ED under the 75% government cap.

Aug 18, 20268 min read

02Funding

SR&ED for Canadian Corporations: Turn Innovation Costs Into a 35% Refund

How Canada's SR&ED program pays back 35% of your innovation costs as a refundable credit, the four things every eligible project must document, and a calculator that estimates your cheque.

Aug 18, 20268 min read

03AI Comparisons

Human-on-the-Loop vs Human-in-the-Loop: Supervise or Approve

Human-on-the-loop supervises a system that acts on its own; human-in-the-loop gates it. Which one a workload deserves comes down to reversibility, not trust.

Aug 17, 20268 min read

04Guides

On-Premise vs Client Cloud vs Managed Cloud: Where Your AI System Should Run

Three places to run an AI system, why data residency selects none of them, and a five-gate quiz that decides between your own hardware, your own cloud account, and a managed environment.

Aug 17, 202610 min read

05AI Explainers

Token-Based vs Scale-to-Zero vs Scale-to-One: How to Run Your Model Layer

Three ways to run inference under your application, the two questions that select between them, and the token volume where a dedicated GPU actually becomes cheaper than paying per token.

Aug 5, 20267 min read

06AI Space

The Best Open-Source Coding Models in 2026 (and How They Compare to Claude)

A practical look at the best open-weight coding models in 2026: Kimi K3, GLM-5.2, Kimi K2.7 Code, GLM-4.7 Flash, and how they compare to Claude Sonnet and Opus.

Jul 17, 202611 min read

07AI Space

Kimi K3 vs Claude Opus 4.8: Open Frontier Meets Proprietary Frontier

Kimi K3 is Moonshot's new open flagship: 2.8T parameters, a 1M-token context, native vision. Claude Opus 4.8 is Anthropic's most capable model. Here's how they compare on capability, cost, and where inference runs.

Jul 17, 20265 min read

08AI Space

Kimi K3 vs GLM-5.2: The Two Open Flagships Compared

Kimi K3 is Moonshot's new 2.8T-parameter flagship with a 1M-token context and native vision; GLM-5.2 is Z.ai's agentic coder and DesignArena's #1 for web and UI. Here's how they compare and which flagship to run for what.

Jul 17, 20265 min read

09AI Space

Kimi K3 vs Kimi K2.7 Code: What Moonshot's New Flagship Changes

Kimi K3 is Moonshot's newest flagship: 2.8T parameters, a 1M-token context, native vision, and dialable reasoning. Here's how it compares to Kimi K2.7 Code on architecture, context, coding, and cost, and when to run each.

Jul 17, 20265 min read

10Perspectives

The money isn't in AI agents. It's in the plumbing.

Gartner says 40% of agentic AI projects will be scrapped by 2027, and the model is never the reason. After a decade shipping AI in regulated operations, here's where the value actually sits.

Jul 17, 20267 min read

11AI Explainers

How to Reduce LLM Hallucination

Hallucination is an incentive and system-design problem, not a mysterious glitch. You reduce it by grounding, constraining, verifying, and measuring, not by waiting for a perfect model.

Jul 14, 202610 min read

12AI Comparisons

Loop Engineering vs Subagent-Driven Execution

Two ways to push an agent past a single pass: keep one context iterating in a loop, or fan the work out to fresh subagents. They optimize for depth versus breadth.

Jul 14, 20269 min read

13AI Space

Migrating from Claude Code to Codex

A practical map for moving from Claude Code to OpenAI's Codex CLI: what the permission flags, headless mode, skills, and config actually translate to, and what you don't need to migrate at all.

Jul 14, 202611 min read

14AI Space

Superpowers vs Opusplan

Superpowers and opusplan both aim at better output from Claude Code, but one is a disciplined workflow and the other routes a smarter model to planning.

Jul 8, 20269 min read

15AI Explainers

What is claude -p?

claude -p runs Claude Code without the interactive terminal: it takes a prompt, does the work, prints the result, and exits, which makes it scriptable.

Jul 8, 20268 min read

16AI Explainers

What is LLM-as-a-judge?

LLM-as-a-judge uses one language model to score another's output against a rubric, scaling quality evaluation that exact-match checks cannot express.

Jun 28, 20269 min read

17AI Explainers

What is an AI agent?

An AI agent is a system that uses a language model to decide and act in a loop, calling tools to accomplish goals rather than only generating text.

Jun 25, 20269 min read

18AI Space

GLM-4.7 Flash vs GLM-5.2: Which Z.ai Model Should You Run?

Both are Z.ai open models, but they sit at opposite ends of the size-cost curve. Here's how GLM-4.7 Flash and GLM-5.2 compare on capability, speed, context, and price, and when each is the right pick.

Jun 22, 20265 min read

19AI Explainers

What is RAG (retrieval-augmented generation)?

RAG retrieves relevant documents at query time and feeds them to a language model, grounding answers in your data instead of pretrained memory.

Jun 22, 20269 min read

20AI Space

How to Run Kimi K2.7 Code in the opencode Desktop App (via AI Space)

Run Kimi K2.7 Code in the opencode desktop app via AI Space using the custom-provider form. Field-by-field setup, model picker, diff review, and fixes.

Jun 19, 20265 min read

21AI Space

How to Run Kimi K2.7 Code in VS Code with the opencode Extension (via AI Space)

Run Kimi K2.7 Code inside VS Code using the opencode extension and AI Space: install the extension, store your key, add a short provider block, and code from the editor.

Jun 19, 20267 min read

22AI Explainers

What is human-in-the-loop AI?

Human-in-the-loop AI keeps a person in control of consequential decisions, letting an agent do the work while a human approves anything irreversible.

Jun 19, 20269 min read

23AI Space

GLM-5.2 vs Claude Sonnet 4.6: Which Is Better for Coding in 2026?

A practical comparison of GLM-5.2 and Claude Sonnet 4.6 for real coding work: capability, workflow fit, cost, and where each model's inference runs.

Jun 18, 20266 min read

24AI Space

How to Run GLM-5.2 in VS Code with the opencode Extension (via AI Space)

Run GLM-5.2 inside VS Code using the opencode extension and AI Space: install the extension, store your key, add a short provider block, and code from the editor.

Jun 18, 20267 min read

25AI Space

How to Run Kimi K2.7 Code in opencode (CLI Setup via AI Space)

Connect opencode to Kimi K2.7 Code through AI Space: store your key with the interactive prompt, add a short provider block, switch models, and fix issues.

Jun 18, 20265 min read

26AI Space

How to Run GLM-5.2 in the opencode Desktop App (via AI Space)

Set up GLM-5.2 in the opencode desktop app through AI Space using the custom-provider form. Field-by-field walkthrough, model picker, and troubleshooting.

Jun 17, 20265 min read

27AI Space

How to Run GLM-5.2 in opencode (CLI Setup via AI Space)

Point opencode at GLM-5.2 through AI Space: store your key with the interactive prompt, add a short provider block, switch models, and troubleshoot.

Jun 16, 20265 min read

28AI Space

Kimi K2.7 Code vs Claude Opus 4.8: A Developer's Comparison

Kimi K2.7 Code is a capable open coding model; Claude Opus 4.8 is Anthropic's most powerful model. Here's how they compare for real development work.

Jun 16, 20266 min read

29AI Explainers

What is an agentic workflow?

An agentic workflow blends deterministic steps with model-driven decisions, giving an AI freedom where judgment is needed and structure everywhere else.

Jun 16, 20269 min read

30AI Space

GLM-5.2 vs Kimi K2.7 Code: Which Open Model Should You Use?

GLM-5.2 and Kimi K2.7 Code are both strong open coding models. Here's a direct comparison on context, speed, agentic tasks, and which fits your workflow.

Jun 14, 20265 min read

31AI Comparisons

CrewAI vs LangGraph

CrewAI and LangGraph both build agent systems, but one models a crew of role-based agents while the other models explicit stateful graphs.

Jun 13, 20269 min read

32AI Space

How to Use GLM-5.2 with Claude Code (Step-by-Step)

Route Claude Code's requests to GLM-5.2 via AI Space in under five minutes. Full setup guide: API key, env vars, model switching, and troubleshooting.

Jun 10, 20265 min read

33AI Comparisons

n8n vs custom AI agents

n8n offers fast visual automation with AI nodes; a custom AI agent trades that speed for control, testability, and production durability.

Jun 10, 20269 min read

34AI Space

Claude Code Alternatives: Running Open Models Without Leaving Your Workflow

You don't have to abandon Claude Code to use open models. Here are three ways to swap the backend (proxy, self-host, or direct API) with an honest tradeoff table.

Jun 8, 20266 min read

35AI Comparisons

OpenAI Agents SDK vs LangGraph

The OpenAI Agents SDK offers a lightweight, fast path to agents; LangGraph offers explicit graph-based control for stateful, durable workflows.

Jun 7, 20269 min read

36AI Space

The Real Cost of Claude Code — and How Open Models Compare

Claude Code billing can be unpredictable at scale. Here's how subscription vs token costs actually stack up, and where open models via AI Space change the math.

Jun 5, 20266 min read

37AI Comparisons

MCP vs A2A

MCP and A2A solve different interoperability problems: one connects agents to context and tools, the other connects agents to agents.

Jun 4, 20268 min read

38Guides

Where Does Your Code Go? A Guide to AI Inference Data Residency

AI inference data residency explained: what it means, why it matters for your prompts and source code, and how to control where inference actually runs.

Jun 3, 20266 min read

39AI Comparisons

LangGraph vs Pydantic AI

LangGraph and Pydantic AI both help build agent systems, but one centers graph orchestration while the other centers typed Python agents.

Jun 1, 20269 min read

40Guides

Data Sovereignty for AI Coding Tools: Why Inference Location Matters

Data sovereignty for AI coding tools explained: what risks your prompts carry, why open weights don't equal local inference, and what sovereign inference looks like.

May 30, 20266 min read

41AI Comparisons

Mastra vs Flue

Mastra and Flue are TypeScript-first agent frameworks, but they emphasize different paths to durable, production-grade AI systems.

May 29, 20269 min read

42Guides

Is It Safe to Send Proprietary Code to Overseas LLM APIs? What Engineering Leaders Should Know

A clear-eyed look at what risks exist when sending code to overseas LLM APIs — jurisdiction, IP exposure, and practical ways to reduce exposure without losing capability.

May 27, 20266 min read

43AI Comparisons

LangGraph vs LangChain

LangGraph and LangChain are related but not interchangeable; one focuses on stateful graph orchestration, the other on LLM app building blocks.

May 26, 20269 min read

44AI Space

Open Models, Western Infrastructure: Frontier Open LLMs Without Off-shoring Your Source

How open-weight models decouple model choice from inference location, and how AI Space runs GLM-5.2 and Kimi K2.7 Code on Cloudflare's Western network for teams with residency requirements.

May 23, 20265 min read

45AI Explainers

What is LangGraph?

LangGraph is a graph-based framework for building stateful, controllable agents and long-running AI workflows.

May 23, 20268 min read

46Guides

GDPR, Data Residency & LLMs: A Practical Guide for Dev Teams

GDPR and LLM data residency explained for developers: lawful basis, EEA transfers, what counts as personal data in prompts, and a practical compliance checklist.

May 20, 20267 min read

47AI Explainers

What is Mastra?

Mastra is a TypeScript agent framework for building AI agents, workflows, memory, evaluations, and production-oriented AI applications.

May 20, 20268 min read

48AI Explainers

What is AG-UI?

AG-UI is a protocol for connecting AI agents to user interfaces through events, state updates, messages, and tool-call streams.

May 17, 20268 min read

49AI Explainers

What is an MCP server?

An MCP server exposes tools, resources, and prompts to AI clients through a standard protocol instead of one-off integrations.

May 14, 20268 min read

50AI Explainers

What are agent tools?

Agent tools are the functions, APIs, data sources, and actions that let AI systems move from conversation to controlled execution.

May 11, 20268 min read

51AI Explainers

What is the A2A protocol?

A practical guide to the Agent2Agent protocol, how A2A lets agents discover capabilities, exchange tasks, and coordinate across systems.

May 8, 20268 min read

52Hiring

Hiring Co-op Students in Ontario: The Career Ready Fast Path

Two stackable programs, the Career Ready Program and Ontario's Co-operative Education Tax Credit, can cover more than half of a co-op student's salary. Here's the worked math.

May 5, 20266 min read

53Academy Video

Host your Lovable app on Cloudflare

Move your Lovable app off the default preview URL and onto Cloudflare, a fast, free, production-ready edge for your custom domain.

May 2, 20266 min read

New posts,the moment they ship.

Drop your email and we'll let you know when new courses and lessons drop. No spam, just the next release.

Loading form…