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.

Layered context cards illustrating an editable agent working set AI & Agents

Context Language Models vs Compaction: What to Pilot

Should your agent edit its own context? Compare CLMs with compaction, inspect the real harness and license, and test what happens when essential facts disappear.

Local Caveman proxy compressing noisy coding-agent context before model inference AI & Agents

Caveman 3.0 for Claude Code: Local Input Compression, Recovery and Benchmarks

Caveman 3.0 puts an Apache-2.0 compression layer between coding agents and models. Here is what its proxy, recovery path, Learn profiler and 33.2% input-token benchmark actually prove.

Wavect editorial illustration for LiteLLM Lens agent trace analysis AI & Agents

LiteLLM Lens: Agent Trace Analysis with SQL and APIs

Move from individual agent traces to evidence-linked investigations. A practical guide to LiteLLM Lens, ClickHouse SQL, safe API access and regression testing.

Wavect editorial illustration for ChatGPT Dots, GitHub bug triage and reviewed fixes Delivery & QA

ChatGPT Dots + GitHub: From Bug Report to Reviewed PR

Turn bug reports into reviewable fixes, not unchecked merges. A practical Dots workflow with a reusable brief, test evidence and clear permission boundaries.

SmythOS Studio self-hosting guide: visual workflows, local infrastructure and runtime boundaries AI & Agents

SmythOS Studio Self-Hosting: Docker Setup, Costs and Limits

An MIT-licensed visual agent builder on your infrastructure. A source-based guide to Docker setup, real costs and the gap between a local demo and a public service.

Two approval paths: review by the Shopify merchant or authorisation by the buyer’s manager Delivery & QA

Shopify B2B Order Approvals: Native Features, Apps, or a Custom Buyer Portal?

Customer approval is not employee purchase approval. Compare native Shopify controls, buyer-team apps and the custom layer needed for budgets and ERP handoffs.

One returns workflow branches into separate inventory and payment controls for a Shopify ERP integration Delivery & QA

Shopify–ERP Returns and Refunds: When the Standard Connector Is Not Enough

One full refund is not an integration test. Use 18 acceptance scenarios to check partial refunds, split invoices, exchanges and recovery before replacing a connector.

Abstract bars representing implementation, operation and migration in a three-year internal-tool cost comparison Business & Regulation

Retool vs Custom Internal Software: A Three-Year Cost Comparison That Includes Maintenance

Three transparent cost scenarios for the same internal application. Include maintenance, user mix and migration without inventing Enterprise pricing or a universal break-even.

Layered paths representing supported file exchange, read-only data access and controlled ERP integration Delivery & QA

Integrating an ERP Without an API: File Exchange, Database Access, RPA, or Replacement?

No API does not mean RPA is the only option. Compare supported access paths, prove recovery, and decide whether to integrate, keep a manual boundary or replace.

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 →