Green tests aren't permission to merge
Every Dependabot PR sitting in your queue is waiting on a person to go verify it. Software Factory Ep. 2 moves that investigation into a sandbox. Plus why software factories aren't cracked yet, and Greg Wilson's Grafana-routed triage agent for incidents.
Coding agents made me more bullish on Rails, not less
This week's video argues coding agents strengthen the case for opinionated frameworks, using a real background-job rebuild as proof. Plus a deep look at why RuboCop, Sorbet, and RubyCritic matter more once an agent is writing the code, and Peter Steinberger's agentic engineering workflow.
How much should you delegate to your AI agent?
This week's video kicks off Software Factory, a new build series with Mastra on scoping agent autonomy one station at a time. Plus Greg Wilson's Grafana dashboard for watching agents in production, a new evidence-driven code review skill, and a 27B-class model that now fits on a phone.
Fractional CTO pricing didn't drop. What you get for it did.
This week's post breaks down what fractional CTO engagements actually cost in 2026, and why the fee held steady while what you get for it changed. Plus my new video on the headcount math that just broke, Greg Wilson's Mastra-to-Prometheus metrics exporter, and a sharp read on career advice in the agent era.
My agent could find the problem. It couldn't fix it.
This week's video: I used ACP, the Agent Client Protocol, to let my Mastra agent hand coding work to Claude Code, with a human approval gate before anything ships. Plus where ACP fits next to MCP and A2A, and the boundary that makes any of it safe.
How much should you delegate to your AI agent?
This week's video covers the shift from one-shot prompts to recurring engineering loops, plus five concrete examples for dependencies, docs, releases, meeting prep, and RFC drift.
Stop giving your agent every tool
This week's video breaks down tool search, context bloat, and the search-load-call pattern for large agent tool surfaces. Plus two related reads on why tool retrieval is becoming a core architecture layer.
Treat your AI agent like a new engineer
This week's video walks through CreatorSignal and shows how repo-local onboarding context helps Codex make better engineering decisions.
When agent loops should become workflows
This week: where an agent loop should end and a workflow should begin, why routing is the unglamorous backbone of inbox triage, and how deterministic guardrails keep model judgment bounded.
The 4 levers that diagnose broken AI agents
This week: a practical framework for diagnosing agent failures, how to turn failed runs into eval cases, and a few resources for building better harnesses.
Approval gates that actually hold
Why 'always confirm before sending' doesn't hold (and what does), plus three more places agent governance is showing up this month: Mastra's runtime layer, OpenAI's on-device PII filter, and Anthropic's Wall Street agent rollout.
Your AI team doesn't need more people. It needs agents.
Stripe merges 1,300 agent-written PRs per week. Coinbase is restructuring around AI-native pods. This week: what 'supported by a fleet of agents' actually means in practice, plus Codex for Claude Code users and OpenAI's $4B deployment bet.
Your AI assistant doesn't need a bigger model. It needs colleagues.
This week: the supervisor + specialists pattern in Mastra, eight AI colleagues on one local model, a one-sentence prompt that turned four manual rebases into a routine, and three links worth your time.
I built Karpathy's LLM knowledge layer into my AI operator (live)
This week's video: live build of the Karpathy/Holmberg knowledge layer wired into Emma, plus why I've been running her on a local Qwen model since picking up a DGX Spark.
Governing AI agents without killing them
The governance post landed this week. Three failure patterns I've either hit or narrowly avoided, what governance-as-code actually looks like in TypeScript, and the scorecard that tells you which layer to fix first.
Your agent reported success. It failed.
This week's video on why agent logs aren't observability, a new doc-sentinel plugin for catching documentation drift, and three links on the AI-at-home use case.
I cut a production prompt by 28% (and the evals that made it safe)
How I used autoresearch to run 65 autonomous optimization iterations on a production agent prompt, plus the eval mental model that made any of it trustworthy.
I gave my AI agent access to my second brain
This week: giving a Mastra AI agent a persistent file system and hybrid search, the grep-over-files pattern, a new doc-audit plugin, and Boris Cherny hidden Claude Code features.
Claude Code Channels: interact with your agent from anywhere
Anthropic shipped Channels for Claude Code. Text Claude from your phone, pipe in webhooks from CI or Cal.com, and let your agent react while you're away. Plus: four patterns that separate agent-ready codebases.
What Claude Code does in your terminal (and when to pause)
Claude Code isn't just for engineers. This week: a terminal guide for non-technical builders, the agent-ready plugin, and Tobi Lutke's Liquid performance PR.
