AI, agents, and the things we build together.
Field notes from running AI agents in production, the wins, the incidents, and the governance patterns that actually ship. Boris writes with help from a small fleet of agents, and posts marked agent-led were run and drafted by agents. It all runs on curate-me.ai.
Everything on this site runs on curate-me.ai. These numbers are real and update every minute.
Everything else
What 1,000 CI Runs Taught Us About Making CI Work for Agents
Most of our pull requests are opened by agents, and CI is their feedback loop. We pulled a month of runs, found where the minutes and the confusion went, and hill-climbed it in four measured rounds. Here's what moved and why.
What Four Real Jobs and $5 Taught Us About Hill-Climbing Jev
We hill-climbed the Jev decision model on four real jobs, scoring each on tuning data and once on held-out data. What moved the scores, what didn't, and the ten lessons we kept.
AI Search Optimization: What Our Blog's SEO Hill Climb Changed
Nine days of SEO work lifted an agent reading test from 26 to 35 and 36 of 40, mostly by fixing our own scanner. It brought no more search visitors, and Google has indexed 28 of our 52 pages.
We Gated Agent Upgrades on a Score and Learned How Weak It Was
A new agent version that scored worse than the best one was thrown away. Then audits of 165 runs found that 46% of reports scoring 85 or higher failed.
Agent Harness vs a Single LLM Call: Why Our $0.875 Job Dropped to $0.0016
The same AI news job cost about 500 times less after we moved it off an agent harness. A controlled replay put the harness itself at about 2x.
We Turned AI Agent Management Into a Game Where Levels Control Real Settings
Our agent dashboard was about 80% reading, so we rebuilt it as a pixel office where levels unlock settings the runner enforces. What it taught us about XP: pay for passed checks, make the checks read the claims, and let real work do the levelling. Our SEO agent reached level 2 by auditing this blog.
What Are Decision Models? What We Learned Putting Jev to Work
Decision models answer typed questions with a probability instead of text. Here's what they are, what we ran on Jev, where it broke, whether upgrading helped, what our own audit found, and what else exists.
LLM-as-a-Judge, Audited: Our Eval Passed 76 Agent Runs That an Audit Failed
We audited 168 reports from our own AI agents against production records. Our LLM eval agreed with the audit less than half the time, and the LLM-as-a-judge that replaced it reached 61.5%, below our 85% bar.
Building a Second Brain for a Family, and Deciding Not to Ship It Yet
A 51-item plan, 43 production PRs in a week, a 14-day canary on one real household, and a quality bar the instrument could not measure. Why HOLD is right.
How We Use curate-me.ai as the Backend for Family Manager
Family Manager has no backend of its own. It is a generic customer org on curate-me.ai, talking to a mobile facade any client could use.
How We Built Family Manager: Architecture, CI/CD, and a Trust Spine
An experienced developer shipping on Apple's platform for the first time, with AI handling the iOS-specific lifting: modules, trust spine, pipeline, and tests.
Family Manager: An AI Family Assistant That Asks First
Family logistics is a stream of small decisions someone has to hold in their head. Family Manager drafts them and waits for your approval.
From Claude Design to a Production Dashboard: What the Handoff Missed
Our 83-page dashboard redesign shipped sample data and the wrong font; a merge broke 14 files. The April account, plus a production handoff worksheet.
Claude Design, With a Brand Bible: Redesigning the curate-me.ai Marketing Page
Three days after redesigning my blog with Claude Design, I ran the pattern again on a marketing page, with a real brand bible uploaded as the design system.
Framework vs Harness vs Gateway: The Layering That Finally Made the Agent Stack Click
We've been talking past each other about three layers of the AI agent stack. Here's the taxonomy that made it click, and where governance actually belongs.
Redesigned My Blog With Claude Design: One URL, Whole-Site Remodel
I redesigned its-boris.com using Claude's new Design tool. Handoff was a single URL. Implementation was one prompt. Here's what I learned, with before/after shots.
Two Weeks Before Launch: Infrastructure, Billing, and a Category Pivot
Two weeks of launch prep: we split our VPS in two, caught a real webhook bug, ran a k6 smoke test, added Microsoft Teams, and pivoted our category.
We Shipped 5 Agent Fleet Templates. One of Them Is a 6-Agent Hospital CFO.
Spinning up a multi-agent fleet from scratch takes two weeks of wiring. We got tired of redoing it, so we shipped prebuilt fleets with governance baked in.
AutoResearch: How We Use AI to Improve Our Own Platform
We built an autonomous experimentation pipeline that runs real experiments on our B2B dashboard: the Karpathy Loop with governance and cost tracking.
Live Build: Ship Code to Production in 3 Minutes
Plain English in, production deploy out: how the 14-stage Live Build pipeline worked, with Claude Code, automated review, screenshots and real numbers.
The Week We Stopped Building Features
82 commits, 45 fixes, 4,700 new tests. After a marathon sprint, we spent three days doing nothing but making the platform actually work.
Marathon Session: 4 Sites, 50+ Features
A full accounting of what got built, broken, and fixed in a single continuous session, from per-org GitHub config to a working AI chat widget.
Why AI Agents Need a Locked Room
135,000 exposed instances. AI agents are powerful, and that's exactly why they need containers, governance, and an audit trail. Here's how we isolate ours.
From GitHub Issue to PR in 3 Minutes: Building an AI Dev Team Pipeline
How a GitHub issue triggers AI decomposition, Claude Code in Docker, automated review and deploy: 179 seconds from issue to PR, and what went wrong.
Week in Review: The AI Dev Team Goes Multi-Repo
An honest retrospective on March 19-24, 2026: 168 platform commits, 11 merged PRs on a new repo, per-org GitHub config, every bug that bit us, and why mounting the Claude Code auth file failed under Docker-in-Docker.
Dogfooding: How We Use Our Own Dashboard to Ship Better
We build a governance platform for AI agents and also run AI agents. How our own cost attribution, error tracking and evaluation framework changed how we ship.
Sprint Retro: What Claude Code and I Actually Shipped (And What Broke)
A real engineering retrospective on 10 days of human+AI pair programming: 319 commits, 3 production bugs, and a billing error that overcharged about 50x.
Why We Switched from MiniMax to StepFun (And Found a 50x Billing Bug)
A model swap that should have been one line became a 46-file investigation, uncovered a billing bug, and showed why model governance needs one source of truth.
How I Built an AI Gateway for OpenClaw
Building a governance gateway for OpenClaw: a base URL swap, a policy chain, cost caps, PII scanning and token-level tracking, with no code changes.
How We Cut AI Agent Costs by 85%
A case study from our fashion analysis pipeline: swapping models, semantic caching and DeepSeek took it from about $0.03 per analysis to under $0.005.
Agent Identity: The Enterprise Blind Spot
Only 21.9% of teams treat AI agents as independent identities. The rest have a compliance gap that regulators will notice.
AGENTS.md, MCP, and A2A: The Three Standards Every Agent Team Needs
The protocol landscape for AI agents has crystallized. Here's what each standard does, how they compose together, and what your team should adopt now.
From Single Agent to Fleet: What Changes
Moving from one AI agent to nine changed five things: who talks to whom, cost allocation, failure cascading, observability and configuration management. Here's what we learned, and what we'd build first next time.
Human-in-the-Loop Patterns for AI Agents
Five human approval patterns we use, from simple approve/reject to iterative feedback loops. How to keep humans in control without slowing agents down.
Perplexity Computer and the Thin Client Pattern
Perplexity's multi-model agent orchestrator validates the thin client approach we've been building. Here's what we learned, and what we shipped.
The Agent Observability Gap
Green dashboards can hide garbage output. The gap between operational health and semantic correctness, and what agent observability actually requires.
The SOUL.md Pattern: Giving Agents Personality
Every agent on this blog has a SOUL.md file that defines its personality, constraints, and communication style, kept separate from its task instructions.
Cost Tracking Across an Agent Fleet
Running a fleet of AI agents gets expensive fast. Here's how we track costs per agent and per model, and how budget caps prevent runaway spending.
Iterative Refinement: How AI Reviews AI
How the blog uses a quality loop where one model writes and another reviews, with configurable thresholds, max iterations, and score tracking.
Agent Webhook HMAC Signing and Retries: What Our Receiver Checks
Sign the exact UTF-8 webhook body, see why reformatted JSON fails, and distinguish a duplicate delivery from completed work. Our receiver contract and an offline Node demo.
Agent Governance: Why It Matters
AI agents are powerful, but without governance, they're unpredictable, expensive, and impossible to audit. These are the controls we put around our own agents, as of March 2026.
How the Agent Fleet Actually Works
A deep dive into the 9 OpenClaw agents that run this blog: their profiles, schedules, tools, and how they coordinate through the curate-me.ai gateway.
Building a Blog With AI Agents
How I rebuilt my blog from scratch using OpenClaw agents governed by curate-me.ai, and why the blog itself is the best reference app for the platform.
Not just blog posts, watch the agents work
Every idea on this site has a runnable version. Click in, watch the terminal stream, see the cost counter tick up.