I missed my own posting rhythm.
My goal is at least one blog post per week. Between July 19 and August 2 I published nothing here.
That gap is not because the work stopped. I got married in Madrid, drove back through France with half the paperwork still in flux, and let the personal week win over the publishing habit.
Fair trade.
Still, the habit matters. This site is my public lab notebook, and silence makes it harder to see whether anything lands.
So I am back with a personal update, not a tutorial. These are notes from the last few weeks of investigation and implementation around AI, plus a few predictions I am willing to put in writing.
I wrote separately about Engram as a shared memory layer and about Secondbrain as a governed assistant. Where does a thought live before it deserves a permanent note?
If every agent writes straight into the vault, you get noise. If nothing is shared, you re-explain context every session. I use a thin memory tier with explicit promotion rules instead of those two. Memory is a product boundary, not a feature.
The agentic design-pattern decision tree post forced me to label workflows instead of defaulting to "ask the big model." Recent weeks I applied that lens to Ayentic support routing (when to suggest vs when to stay silent), PublicBrain publish gates (deterministic checks before any model touches content), and personal cron jobs (rules first, LLM only on exceptions).
Teams that skip this step pay in token bills and incident postmortems.
Model Context Protocol showed up everywhere in my toolchain: Prismic, browser tools, analytics hooks, memory.
The failure mode is the same as microservices in 2016. Too many endpoints, no ownership. I treat it as a wiring layer, not magic. I am standardizing on a small set of servers per project and documenting which agent may call which tool.
Default deny. No model call until a cheaper path failed.
Tiered models. Small or local for classification and drafts. Frontier only for synthesis or ambiguous cases (cost notes).
Audit tags per request, especially for anything client-facing (Sureva, Ayentic).
A human publish gate for public text. That is this site's PublicBrain pipeline.
Measure, or admit blindness. Until this week I had no site analytics. We just added Google Tag Manager so GA4 and future pixels can ship without redeploying the app.
That last point is embarrassing for someone who preaches pipelines. The fix was one env var plus a layout hook. Habit applies to infra too.
Agent memory becomes a SKU. Vendors will sell "memory APIs" the way they sold auth. Winners will expose export, delete, and scope clearly. Swiss and EU buyers will ask.
Most AI products collapse to three UI modes: copilot sidebar, inbox with suggestions, batch automation. Ayentic is explicitly inbox-first. I expect convergent evolution everywhere else.
Publishing cadence beats model upgrades for solo builders. A steady public log compounds trust and SEO faster than swapping GPT-N every quarter. I am testing that on myself.
On-prem inference stops being a niche. Municipal and SME buyers in CH/DE want deployment stories, not slide decks. Sureva's lane and my barto-side prototypes point the same direction from different angles.
Measurement returns as a discipline. After a year of "ship the demo," teams will ask which pages and prompts actually convert. GTM and GA4 are table stakes. The interesting part is tying blog posts to inbound calls.
Multi-agent hype cools. Orchestration stays. Fewer swarms, more explicit graphs with retries, budgets, and kill switches. Closer to workflow engines than sci-fi.
I am already back to one post per week minimum. This is week one. GTM is live on johnkegd.com so I can see if anyone reads the CRIF or Madrid posts. The next deep retrospective IT post I already named is the Ajila to Swisscom bridge, then Solmedic and barto. On Engram and PublicBrain I am tightening promotion from session memory to vault drafts. Still no auto publish.
The wedding post was life. This one is work returning to the surface.
If you are building agents in production, the honest metric is not which model. It is whether the system survives contact with reality for a month without you babysitting it.
That is the bar I am holding myself to this summer.