FROM THE TRENCHES Issue №354

Practical decisions for AI and software

Technical guides to choose models, verify software and plan delivery.

Latest articles

New comparisons, practical guides and engineering lessons.

One English slug per document across four language prefixes, with hreflang links pointing between the versions AI & Agents

Localized URL Slugs vs hreflang: Why We Keep One English Slug

Localized slugs are valid, but they add a mapping layer around hreflang. Why we keep one language-independent slug across four languages, and what that costs us.

An Odoo database beside a team's own integration service, with the call ceiling, the transaction boundary and the upgrade clock drawn between them Delivery & QA

Odoo API integration: the five limits that decide your architecture

Odoo Cloud has a call policy, each JSON-2 call has its own transaction, and legacy RPC services retire in stages.

An agent reading a page on one side and completing a task against an authorized tool surface on the other AI & Agents

Can an AI Agent Use Your Product, or Only Read About It?

Readable means an assistant can cite you. Usable means it can complete a task and be refused when it should be. The second is an authorization problem, not a model problem.

A coding agent repo map split into a committed team artifact and a regenerable local cache AI & Agents

Graft Review 2026: Do Agent Repo Maps Belong in Git?

Graft writes your codebase into linked markdown so agents stop re-exploring it. We checked whether that map really travels through git, what its benchmarks prove and what belongs in version control instead.

A prompt passing through a pseudonymization layer, with the re-identification key staying on the controller side Business & Regulation

LLM Pseudonymization Gateways: Does the Prompt Leave GDPR Scope?

Platforms in this category promise that pseudonymizing a prompt makes a hosted model safe to use. The 2025 CJEU ruling and the EDPB guidance say something narrower. Here is the evidence a buyer actually has to collect.

Developer terminal session with compressed tool output AI & Agents

Codag for Coding Agents: Tool-Output Compression and Cost

Coding agents can waste many tokens on oversized tool output. Learn how output compression and decision-first observability make AI execution costs measurable again.

Lean token budgeting for AI coding agents AI & Agents

Smarter Token Usage: Measure Cost per Accepted Coding Task

Reduce coding-agent spend with cache-aware accounting, tested routing and a reproducible cost-per-accepted-task example, including retries and human review.

DeepSeek Harness plugin layers connecting models, tools, permissions and the agent loop AI & Agents

DeepSeek Harness Review: Is the Plugin Stack Production-Ready?

A buyer-focused review of DeepSeek Harness, its replaceable Cordis architecture, four agent presets, security boundaries and enterprise pilot fit.

Layered vLLM and NVIDIA Triton inference stack AI & Agents

Netflix’s vLLM and Triton Stack: 7 Production Lessons

Netflix chose vLLM for operational fit, then found the real bottlenecks in constrained decoding, deployment, model loading and metrics. Here is what smaller teams should copy, and what they should buy instead.

Inbox, without the noise

Get the next field note

One concise email when we publish. No tracking pixels, and no inbox filler.

Facing one of these decisions right now? Talk it through with us →