Notes from the working lab.
Practical accounts of building AI workflows with visible boundaries, human decisions and evidence that survives contact with reality.
I Gave My AI Browser a Recovery Routine
My AI team and I repaired an assistant's visible browser connection and tested recovery from a stopped screen. The instructions now define what it can restart, what it must verify and when a failure needs a different response.
My Weekly AI Review Needed My Ranking, Not Just a Summary
My AI team and I sent the first weekly review email. My ranking feedback made the priorities concrete, while a corrected archive remained distinct from the original delivered edition.
I Built a Maths Practice Site with My AI Team
My AI team and I built a maths practice site with generated exercises, answer checks, hints and worked solutions. It grew from a spatial Pythagoras prototype into two chapters of practice.
I Moved a Work Routine, Not Just an AI Prompt
My AI team and I moved Work Focus to my coaching assistant, including its historical journals and scheduled expiry job. The handoff involved a running routine, not just a new instruction.
My AI Wiki Hasn't Earned Automatic Upkeep
My AI team and I completed a linked Obsidian wiki experiment. I judged it not yet useful enough for automatic upkeep: a finished build and a worthwhile recurring workflow are different outcomes.
My AI Setup Needs More Than a Mirror
A versioned off-host backup layer replaces a one-off mirror, with database-aware backups and recorded recovery tests. The full disaster-recovery rehearsal remains a separate milestone.
My Recording Notes Need the Recording Date
We updated recording filenames, the web interface and exports to carry recording time, while preserving originals and processing results.
My Shared Goals Need More Than a Better Screen
My AI team and I connected recorded goals and actions to a daily briefing, an e-ink view and a weekly review. A better display is only one part of the workflow.
My AI Memory Needs a Language Check Before It Saves
My AI team and I repaired unexpected language in stored memories and added a pre-storage safeguard with a corrective retry and rejection if the check fails again.
An AI Audit Needs to Correct Its Own Findings
My AI team and I updated five Python dependencies and withdrew a backup warning caused by a restricted filesystem view. The audit needed correction as well as remediation.
A Private API Is Not Yet a Working Phone Connection
My AI team and I tested a private API route for phone access. Authentication worked on the server, but the iPhone connection is still an open test.
I Want My AI Work to Sound Like Me
I want a daily blog about building with my AI team, written from my perspective and shaped by feedback rather than turned into another work log.
Who Authorized the Work Is Not Where It Ran
Why consequential agent workflows need separate evidence for workload identity, execution environment and human business acceptance.
An Agent Directory Needs an Owner, Not Just a Listing
A reusable agent deserves more than a searchable name. Five visible signals can help people judge whether it is current, appropriately scoped and worth considering for reuse.
When a Provider Rule Changes the Planning Mode
A group plan needs to show the provider rule that changes the next step, whether that means a different fare path, a timed slot or a human-owned check.
Plan Something Everyone Can Join in Their Own Way
Good group coordination does not require equal real-time participation. A better invitation gives people clear ways to contribute, observe and confirm at the level that suits them.
Make Multi-Agent Stage Boundaries Explicit and Restartable
A recoverable agent workflow needs more than a completion message: each stage should declare its dependencies, prove its artifact, and specify when the next stage may begin or the plan must be restarted.
The PatOS AI factory starts with an access path we can trust
How the first PatOS AI factory step established a human-controlled route into a live Lab VPS while keeping the SSH private key in Bitwarden.
When a Group Plan Changes, the Next Step Should Be Clear
A shared plan is not finished when people agree to it. It has to keep serving the people affected when a time, place, ride or reservation changes.
A Group Chat Is Not a Commitment: Make the Plan Confirmable
A message, link, or booking receipt can be useful evidence. It should not become a consequential group decision until the right person can see it, understand it, and confirm it.
When an AI Recommendation Changes, You Need More Than a New Answer
AI-assisted planning becomes easier to trust when a changed recommendation can be traced to the source, the constraint, or the execution path that changed.
Before You Add an Agent, Pass the Architecture Gate
Multi-agent work is an operating mode to test, not a default upgrade. Start with a single-agent baseline, classify the task, choose the simplest topology that fits, and make every handoff reviewable before you scale collaboration.
Before You Add AI to the Team, Decide Where It Fits
The useful first question is not whether AI can do a task. It is which existing role and recurring decision it can improve without losing the context, judgment and accountability that make the work trustworthy.
The Map That Remembers Where You’ve Been and Finds Somewhere New
How Rides turns personal e-bike history and open map data into inspectable routes that favour roads you have not ridden yet.
Cloudflare MCP Without Handing an Agent the Keys
Why PatOS uses Cloudflare MCP for bounded discovery and preparation while keeping consequential authority human-owned.
Secure by Design: From Ambient Secrets to Scoped Access at PatOS
How scoped machine access and deliberate authority boundaries reduce ambient secret exposure in agent-assisted workflows.