The router mess is a design lesson: own your tool-calling layer
Outages, throttling, and shady middlemen keep breaking LLM apps. A look at what to own yourself, and how to build a thin, honest tool-calling backend on VicroCode.
AI Market Guide
Evidence-based AI market signals, coding demand, and practical growth cases for builders.
Practical guides
Don't Trust the Agent's Auto Mode: Building an Editable Approval Gate With a SQLite Decision Log
A build debrief on replacing an agent's built-in auto-approval with a small policy engine that classifies each tool call, logs every decision, and holds risky ones for a human.
Give your coding agent a profile you can edit, not one it quietly infers
A V2EX thread showed an agent auto-summarizing a developer's habits into a profile. Here's a build debrief on making that profile a human-owned artifact instead.
Ditching context compaction: an agent that keeps its working memory in editable SQLite records
A build debrief on swapping session-history compaction for a queryable context store: live SQLite records the agent revises via transactions, retired items pushed to a vector store.
Making an Agent's Wait Stop Being a Black Box: A Live Tool-Call Status Board
A dev wished he could see which tool his model was calling during long waits. Here's a reproducible Python + hosted HTML board that streams thinking vs tool state.
Building a Token Cost Tracker Before You Commit to an Agent Harness
Caijing's Aug 2024 test showed domestic subscriptions burning quotas in days. Build a Python monitor that logs cache rates and per-task burn before choosing a harness.
Don't Trust the Spec Sheet: Building a Token-Frugal Agent That Logs Its Own Tool Calls
Meta says Muse Spark 1.3 cuts tool calls ~20% and tokens ~25% (vendor figures, unverified). Here's how I turned that promise into something I could actually measure on VicroCode.