The half of the workshop
where the robots work.
I run a crew of AI agents with real tool access: a local inference rig that routes its own traffic, a bench of Flipper Zeros, ESP32s, and 3D printers wired into the loop. This site is what they build with me, and for me. It is also where they answer for it. Logged daily, with receipts.
System status
What the lab ships
Six lanes of work, each with a crew member on it. The robots are the workforce; I just hold the soldering iron.
AI agents
Nexus on duty
A roster of governed agents with eval gates, audit trails, and memory with provenance. They triage, summarize, investigate, and file their own receipts.
Local inference
Forge on duty
A Strix Halo rig running a 190+ model fleet behind a routing gateway. Chat, embeddings, vision, voice, and image generation without leaving the building.
Maker bench
Builder on duty
Flipper Zero, ESP32s, breadboards, a 3D print pipeline with its own MCP server, and wearable capture experiments. Solder smoke included.
Automation
Operator on duty
Self-chaining job schedulers, daily blog generation, deploy gates, and the boring glue that keeps a hundred repos honest.
Security tooling
Sentinel on duty
DB-isolated security analysts, read-only eval contracts, and a healthy distrust of anything that can't show its logs.
The build log
Scribe on duty
Every day the pipeline drafts what got shipped, a human approves it, and it lands here. The receipts of the whole operation.
Build log
The Daily Builder's Blog
Drafted by the pipeline from real work, reviewed by a human, shipped here. What actually happened, daily-ish.
Optimizing Local Inference: Solving the 'Thinking' Budget Issue
How disabling reasoning flags fixed empty responses in local Gemma 4 12B deployments.
Refining the Worker Suite and Discord Delivery
Fixing parser assertions and improving embed aesthetics across the ecosystem.
Retrieval Latency, Deploy Gates, and Agent Governance
A day of fixing cold search performance, unblocking MCP deploys, and hardening agent runtime governance.
Notes / Doctrine
How the crew is kept honest
Agents with tool access are a security problem wearing a productivity costume. Twenty years in security taught me to treat mine accordingly.
Candidate-first
Agents propose; the operator approves. Nothing gets promoted to durable truth or production without a human decision on the record.
Explicit tool grants
Every agent runs against a declared grant profile. If live grants drift from the manifest, certification blocks until they reconcile.
Eval gates
Read-only contracts proven by deterministic test fixtures before an agent activates. The security analyst passed its isolation suite before it touched anything.
Receipts
Every LLM-backed decision is traced. Agents produce claims; git and the warehouse hold the evidence. If it is not observable, it did not happen.
SENTINEL DOCTRINE / TRUST, BUT VERIFY THE LOGS
Details / Field notes
Dispatches on DEV
Longer-form writeups from the workshop, published on dev.to. Pulled from the feed, refreshed hourly.
My agent swarm had a productive night. My pipeline lied about it.
I gave a Grok CLI agent swarm one instruction around 1am and went to bed. By 5:30 it had closed 37 items off my operator backlog and landed 20 commits on main: 2,838 lines added...
READ ON DEV.TO ↗GitHub Got Breached Through a VS Code Extension. MCP Servers Are Next.
Yesterday, GitHub said it had detected and contained a compromise of an employee device involving a poisoned VS Code extension. The company said its current assessment is that t...
READ ON DEV.TO ↗Capture the Reasoning Path, Not the Final State
Two files, one discipline, and a measured 10-13% of my Claude Code budget. A while back, mid-session with Claude Code, I typed a pushback in the kind of broken English you only...
READ ON DEV.TO ↗Codex Chronicle was paying for every frame.
I built a four-sensor Gemma 4 replacement on a Mac mini. For about a week I had OpenAI’s research-preview Chronicle running on my MacBook. Every ten minutes it screenshotted my...
READ ON DEV.TO ↗Notes / Longwave
Essays on Substack
The slower frequency: essays and longer thinking, delivered by newsletter. Pulled from the feed, refreshed hourly.
The System Underneath
What ADHD and autism actually feel like when you are doing technical work, and what happened when I built a tool to measure it.
READ ON SUBSTACK ↗People Will Always Find a Way
What desire paths teach us about AI and enterprise software
READ ON SUBSTACK ↗I Built the Operating System Sam Altman Thinks Only College Kids Use
Sam Altman told a Sequoia Capital audience that older people use ChatGPT like a Google replacement, people in their 20s and 30s use it as a life advisor, and college students us...
READ ON SUBSTACK ↗Your iPhone Knows Who Matters to You.
And where you go together, and what they look like.
READ ON SUBSTACK ↗Expressions / Poses
The crew
Make it work. Make it better.
Observe. Record. Explain.
Trust, but verify the logs.
Route the work to the right machine.
Context with provenance.
Go see. Bring back signal.
Somebody has to clean the weird corners.
Notes
What I build myself lives at niclydon.io
Projects, ideas, and the human half of the operation.