I don't need an agent to invent a new way to run my tests or triage a PR every time. Extracting those procedures into code gives me reusable tools, smaller agent jobs, and results I can inspect. It can also give me a very elaborate way to waste time.
I had already done much of the thinking. AI helped assemble the supporting material, take away the blank page, and turn my teaching notes into a handout.
The Software Factory became a bottleneck. I’m scaling it back toward working drafts, bounded agent work, and verification that helps me finish rather than making every iteration start over.
I shut down Hermes and launched Moment Savor on the same day. The connection is a simpler division of labor: Amp handles the novel work, while Swamp makes the repeatable parts observable and deterministic.
Shutterfly timed out after an hour of selecting family photos. The error said to start over. The browser DOM said otherwise.
My pipeline built pull requests one increment at a time because my old laptop couldn't hold two checkouts of a big codebase at once. I replaced that laptop months ago and never noticed the reason was gone — until a coworker asked a simple question on a work trip.
My Valetudo robot was already speaking MQTT. The useful part was turning that topic tree into a discoverable, validated Swamp model instead of making an agent guess.
A fleet-wide upgrade chore turned into a published extension, and its first live campaign caught a bug in itself. Field notes on adopting swamp's new assert steps — and on making release adoption deterministic instead of a vibe.
I exhausted a Claude allowance with an unbounded benchmark, then spent a weekend finding out what a supervised software factory could actually deliver. Forty-nine accepted runs later, the evidence points to controlled delivery—not autonomy.
A Swamp model that turns a few trip details into a categorized packing list, previews before it touches Todoist, and asks before it commits.
I wanted Hermes to stop improvising and start reusing the models and workflows we had already built. The setup was short. Proving it worked took wrong homes, hidden errors, a broken registry, and one extension Hermes finally published and ran through Discord.
Tracking meal expenses on a work trip by chatting with an agent from an elevator, and why the ledger lives in a Swamp model instead of a spreadsheet or a note.
Twenty cameras can see a lot, but the useful footage is scattered across views right when I don't have time to click through a timeline. Swamp lets an agent take the first export pass so review can happen later, from a real screen.
UniFi update notifications are satisfying to clear—until the hardware you depend on turns an ordinary firmware update into a troubleshooting project. Swamp gave mine a second opinion first.
A two-sentence Discord prompt at a rest stop became a route-risk watch, then a quiet home safety system — and a lesson in keeping an agent from over-publishing what it builds.
Two weeks of orchestrated agent swarms, a datastore mess of my own making, and an upstream bug loop that actually worked — the projects I'd stopped starting are moving again.
A three-part series about truth, allegiance, and wisdom while using AI as Christians.
A benchmark can show a model is fast. A structured agent record can show whether it actually earns a place in the workflow.
AI can assist with many kinds of work, but it cannot take the place of Christian wisdom, conscience, responsibility, prayer, or worship.
Christians should take AI’s moral pressures seriously without confusing ordinary use of a tool with worshipful allegiance.
Benchmarks can show us something real, but they cannot carry the weight of prophecy; the truer test is what a model does in the work in front of us.
I did not want a dashboard on my wrist. I wanted a short, trustworthy status signal—and a little progress marker.
I spent a few minutes prompting Hermes during a homeschool conference, and the automation made me more engaged, not less.
How I replaced a manual checklist and AI browser agent with a Swamp extension model that gates PRs with structured, queryable test results — no cloud CI spend required.
Local LLMs on the M5 Max have become useful for bounded agent work, but they are not Amp or Claude killers yet.
Hermes and Swamp finally worked once I stopped chasing install ownership and fixed the session snapshot that every tool subprocess sources.
A Frigate storage alert became a small example of why agents and Swamp are working so well in my homelab: bounded context, reviewable changes, and asynchronous fixes that preserve headspace.
I built an offline meeting transcriber using local Whisper, pyannote, and Ollama, then migrated it to swamp — the same typed models now power my terminal, my agent, and a future watch-folder, without copy-pasting a script.
I built an AI pipeline that ships production code to a ten-year-old Rails monolith about 90% of the time. But when it fails, or when I change something and want to know if it helped, I'm guessing. Three days of migrating to swamp gave me structured data for every decision the pipeline makes.
A pastoral and practical take on AI from a Christian perspective — wise, winsome, and well-timed.
Two things resonated from Simon Willison's post on agent-written code: owning the output regardless of who wrote it, and working code — not artisanal craft — as the real deliverable.
Claude Code handled complex tasks well in plan mode, but using it through a cooldown left me with quick solutions and a shaky mental model — until a post on Naur crystallized why.
The Southwest Companion Pass is often cited as the best deal in travel. But is it worth the effort for most people?
How I've earned over $100 in Google Play Store credits just by answering short surveys on my phone.
A quick tip on how to maximize your Southwest Rapid Rewards points with minimal effort.
Stop overpaying for your cell phone plan. There are plenty of MVNOs that offer the same service for a fraction of the cost.
What is Meagerfindings and why did I start this blog?
Bank account bonuses are one of the easiest ways to make extra money if you have some spare cash sitting around.
A 28-mile loop through the Maroon Bells-Snowmass Wilderness, crossing four 12,000+ foot passes.