Join The Modern Engineer Newsletter

Design a markdown-first documentation context package with llms.txt, source ownership, freshness rules, exclusions, and eval questions so agents can use docs without treating the whole site as context.

Loop engineering moves AI work from one prompt at a time to bounded control systems with triggers, gates, state, budgets, and escalation paths.

JSON, XML, YAML, TOML, and TOON each carry different signals inside a prompt. Knowing what the model actually does with each format is how you make format a deliberate engineering decision instead of an afterthought.

Define a production prompt and context lineage contract that records effective model inputs, protects sensitive content, and supports evidence-based incident replay.

Build a local C# approval workflow where risky AI-proposed actions create explicit approval requests, consume scoped expiring tokens, and leave behind auditable decision records.

Treat long-running agent context as a controlled system. Pin required facts, compact old history, expire stale notes, and carry forward only what the next turn is allowed to rely on.

Treat live MCP servers as versioned tool contracts. Discover live schemas, replay frozen probes, and block risky drift before promotion.

Treat model swaps as controlled experiments: frozen evals, shadow replay, deterministic scoring, and explicit promotion gates before an AI system changes models.

Repository instruction files are becoming the control plane for AI-assisted coding: a short AGENTS.md map, deep docs as the system of record, and task-specific specs for execution.

A gateway that discovers, validates, and delegates to a remote specialist over A2A. Both agents run on Microsoft Foundry and deploy to Azure Container Apps from one image.