Independent AI Software Studio · New Zealand
PapaCoder
Building AI-powered apps and developer media.
Latest article
Evaluate AI Coding Agents at the Patch, Not the Promise
A practical system for evaluating AI coding-agent changes with task contracts, fail-to-pass tests, pass-to-pass regressions, and diff review.
More from the blog
See all →Four boundaries to set before handing a repository to an AI agent
A practical guide to scope, context, authority, and verification when using AI agents on a code repository.
Don't Kill Auto From the Email Alone — Cursor's Aug 24 Per-Model Pricing
From 24 Aug 2026, Cursor Auto bills at the routed model's list price instead of a flat Auto rate. Included limits rose; routing did not change — check Spending before you kill Auto.
Don't Leave JetBrains Outside the Policy — Copilot Enterprise Managed Settings
GitHub Copilot for JetBrains now honors enterprise managed settings (2026-08-18). Close the IntelliJ gap for MCP allowlists, plugins, OpenTelemetry, and Bypass/Autopilot.
Don't Run That headersHelper Blind — Claude Code 2.1.238 Moves the MCP Trust Boundary
Starting in Claude Code 2.1.238, project and plugin headersHelper commands require folder trust and strip credential-like env vars. Treat shared .mcp.json as a shell execution surface — with a practical checklist.
Before You @GitHub in Standup Chat — Teams Shared Agents and the Trust Boundary
Teams Copilot shared sessions (2026-08-21 preview): write-access triggers, sandbox + AI credit budgets, extra PR approval for the Teams identity — companion to Slack Code, not a rehash.
Before @GitHub Ships From Slack — Code Channels and the Trust Boundary
GitHub’s 2026-08-21 Slack agentic preview. Lock write access, app-identity PRs, extra approvals, and cloud-agent budgets before the first @mention.
When the Agent Wakes Itself — Treat Cursor Subscriptions and /goal as a Harness
Cursor’s 2026-08-19 cloud agent release: Subscriptions, /goal, and isolated subagent VMs. How to run always-on loops with stop conditions—not autopilot.