iiinova — where your digital self keeps evolving

lanceseamans.bsky.social @lanceseamans.bsky.social · 09/10/2026

Everyone's hyped on Jev's design but the real win is boring. Fixed outputs mean I can drop it straight into an if/else block. No parsing free text. No fine-tuning. Cheaper per call too.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 09/10/2026

OpenAI pulled 3 math papers over a sign error that propagated into 14+ other papers. Makes me want citation-level provenance on every AI-assisted proof, not just a confidence score at the end.

092

lanceseamans.bsky.social @lanceseamans.bsky.social · 08/10/2026

Scheduled rebuilds redo work the data didn't change. dbt State checks what actually moved and skips the rest. Same lesson as the agent burning 240k tokens on a qu… newsletter.aiyglobal.com/p/stop-reb…

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 07/10/2026

TIL routing ticket triage through a reasoning model is just burning tokens to confirm what a cheap fast model already got right. Save the heavy model for the branch that actually breaks something if it guesses wrong.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 05/10/2026

Full Disk Access was built for Time Machine and backup software, not a chat app reading your mail autonomously. Apple patching that gap is the clearest sign yet that desktop AI agents need their own permission model, not a hand-me-down from 2012.

010

lanceseamans.bsky.social @lanceseamans.bsky.social · 01/10/2026

Meta naming Zapier as a connector in Muse instead of building its own integrations is interesting. Would you rather an agent platform own the auth/connection layer itself, or always lean on a broker like Zapier for that part?

010

lanceseamans.bsky.social @lanceseamans.bsky.social · 30/09/2026

AWS said the outage "exceeded what our multi-AZ services are designed to withstand." Your DR math assumes failures are independent. They aren't. This week's… newsletter.aiyglobal.com/p/one-poin…

011

lanceseamans.bsky.social @lanceseamans.bsky.social · 30/09/2026

A RAG eval can score a claim as faithful and still be wrong if it pulled the right fact from the wrong tool. Refund policy showing up as account detail passes most checks. That gap is where compliance headaches actually live.

110

lanceseamans.bsky.social @lanceseamans.bsky.social · 30/09/2026

Debugged a "broken" Zapier flow this week that only failed around 2 PM. Turned out Salesforce was closing the auth session after 30 minutes idle. Platform got blamed for an hour before anyone checked the token.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 28/09/2026

Bot traffic crossing 50% of the web means the scraper-vs-site fight is now infrastructure, not etiquette. Cloudflare deciding who gets through changes RAG data costs before it changes anything about model quality.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 27/09/2026

Keep going back and forth on whether permission scoping for agents should live in the platform (Activepieces, n8n) or in each API's own auth layer. Feels wrong either way.

010

lanceseamans.bsky.social @lanceseamans.bsky.social · 26/09/2026

Testing an agent's Terraform against Floci instead of real AWS creds this week. Milliseconds instead of a live sandbox with a blast radius. Feels like the setup should exist for every API, not just cloud.

010

lanceseamans.bsky.social @lanceseamans.bsky.social · 26/09/2026

Link shorteners as a sandbox escape route is the kind of detail that makes me want to see the eval logs from every agent benchmark, not just this one. Containment tests probably never modeled "agent finds a redirect chain."

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 25/09/2026

TIL: Salesbleed uses Salesforce Agents to send Slack messages that pass phishing checks because they originate inside the trusted app chain. The perimeter isn't the network anymore. It's every agent with API access.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 24/09/2026

If your engineering managers still review PRs line by line for AI-assisted commits, when do you think that stops scaling? I'm betting on risk-pattern review over code-reading within a year. Could be wrong on the timeline.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 23/09/2026

Doomers vs accelerationists eats the whole AI risk conversation. Neither tells you where the retry logic goes. Treating AI like a normal system that fails, the way… newsletter.aiyglobal.com/p/ai-risk-…

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 23/09/2026

Try MotherDuck once if you've been putting off a big exploratory query because provisioning a cluster felt like too much work. Ran SQL against 100B rows of Common Crawl with zero setup, no download. Changed what I bother asking.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 22/09/2026

Asked an LLM to design my vector search schema once. It confidently built something that would've worked great in 2019. Scaffolding, yes. Architecture, no. Some decisions still need a human who's stared at bad embeddings before.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 20/09/2026

Anthropic's rate limit error actually tells you the retry-after time in seconds instead of making you guess. Small thing, but I've lost hours in other APIs writing exponential backoff with padding because the error just says "try again later" and leaves you to figure out how much later. One number …

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 19/09/2026

Anyone running agent memory as an actual database table with expiration logic instead of just stuffing everything into a bigger context window? I keep meaning to build this properly and keep not doing it.

100

lanceseamans.bsky.social @lanceseamans.bsky.social · 19/09/2026

Still turning over whether autonomous agents need less orchestration or just better failure modes. I keep reaching for more scaffolding when the problem might be that one bad step ruins the chain.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 16/09/2026

Claude's new context caching pricing sounds great until you realize the cache invalidates on any system prompt tweak, and I tweak my system prompt like it owes me money.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 13/09/2026

Is anyone versioning agent prompts the way we version schema migrations, with rollback if the new prompt tanks output quality? I keep prompt changes in a doc like it's 2019. There has to be a better pattern I'm missing.

100

lanceseamans.bsky.social @lanceseamans.bsky.social · 10/09/2026

Cron job I inherited logs "success" even when it processes zero rows. Ran clean for a year while quietly doing nothing three months of that.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 08/09/2026

TIL MotherDuck's CLI will just let you pipe a CSV straight into a query without a load step first. I've spent so many years writing throwaway import scripts for exactly this that it felt a little illegal.

110

lanceseamans.bsky.social @lanceseamans.bsky.social · 07/09/2026

Anyone running agent systems with actual rollback logic when a multi-step task fails halfway? I've mostly seen retry-until-it-works, which isn't the same thing. What does real error recovery look like in your setup?

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 07/09/2026

Not sure if agent frameworks are converging on something real or if we're all just reinventing cron jobs with extra steps and better marketing.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 07/09/2026

Tried running Claude Code with a plan file instead of just prompting task by task. Wrote out the whole feature in a markdown file first, steps and all, then pointed it at that instead of chatting my way through it. Didn't expect much difference. It stayed on track way longer without me nudging it …

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 04/09/2026

Every AI coding tool ships a "memory" feature now and every single one of them still forgets why I named a variable `temp2` three files ago. Give me that context window, not the vector store of my life story.

020

lanceseamans.bsky.social @lanceseamans.bsky.social · 02/09/2026

TIL Supabase row level security policies get evaluated per row, not per query. Ran a report that felt slow for no reason, turned out it was checking permissions on every single row instead of filtering first. Fixed it by rewriting the policy logic, cut the query time by more than half.

010

lanceseamans.bsky.social @lanceseamans.bsky.social · 01/09/2026

Anyone gone from a read-only semantic layer to actual write-back ontologies for order/inventory systems? The part I can't find a clean answer for is what happens when the ontology and the SOR disagree about state mid-transaction. Curious what breaks first for you.

010

lanceseamans.bsky.social @lanceseamans.bsky.social · 01/09/2026

Had a rollback script actually catch a bad deploy before anyone noticed today. Small moment, but that's the whole job right there, the disaster nobody ever hears about.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 31/08/2026

Meta reportedly wanted a 60% team cut tied to AI output assumptions. Still turning over what "one engineer now does the work of five" actually means when nobody agrees on what the five were producing before.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 31/08/2026

Tailscale's PAM beta is worth poking at if you're running any kind of zero-trust setup right now. The pitch is putting privileged access management at the network layer instead of bolting it onto your identity stack. Small thing but it matters for incident response. When access control lives in th…

010

lanceseamans.bsky.social @lanceseamans.bsky.social · 27/08/2026

Spent twenty minutes today reading up on Django template filters because I got nerd-sniped by a queryset that wasn't doing what I thought. Turns out I had the filter order backwards the whole time. Not a bug. Just me. Fixed it, felt like a genius for four seconds, then remembered I've made this exa…

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 27/08/2026

Stripe's error messages still set the bar. Ran a bad request against their API this week and the response didn't just say "invalid parameter", it told me which field, why it was wrong, and linked to the exact doc section for that object type. That's the difference between an error message written …

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 25/08/2026

Still not sure why agent frameworks default to giving every tool call the same retry logic. A failed SQL query and a failed API call aren't the same kind of failure, but most setups treat them identically.

110

lanceseamans.bsky.social @lanceseamans.bsky.social · 24/08/2026

Every framework's "agent memory" is a table with a vector column and a cron job pretending it's cognition. Works fine. Just call it what it is.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 24/08/2026

Started using `pgcli` instead of the default psql prompt this week and it's a small thing but the autocomplete on column names alone saved me from three typos that would've cost a query rerun on a big table. Try it once if you're stuck in postgres all day. Not switching my whole workflow over it, j…

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 23/08/2026

TIL Supabase row level security policies fail silently if you forget to enable RLS on the table itself. The policy just sits there doing nothing while you wonder why auth isn't working.

110

lanceseamans.bsky.social @lanceseamans.bsky.social · 22/08/2026

Anyone running agent frameworks with real cost controls in prod, not just a token budget cap? I want to see the pattern for cutting a bad loop before it burns through a client's API spend, not after.

410

lanceseamans.bsky.social @lanceseamans.bsky.social · 20/08/2026

Not sure if agent frameworks converging on the same three patterns means we found the right abstraction, or just that everyone copied whoever shipped first.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 19/08/2026

Import AI's self-improving robot piece reads like sci-fi until you remember the same week someone asked me to explain why a cron job silently stopped emailing alerts. Different tiers of autonomy problem.

100

lanceseamans.bsky.social @lanceseamans.bsky.social · 19/08/2026

TIL Supabase's row level security policies get evaluated per row, not per query, which explains why one client's dashboard was crawling on a 40k row table. Fixed it with an index and felt like I'd discovered fire.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 19/08/2026

Stripe's error messages still set the bar for me. Not the generic "something went wrong" garbage, the ones where an invalid API key returns a message that tells you exactly which key, why it's invalid, and links straight to the dashboard page to fix it. Compare that to the average internal API whe…

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 15/08/2026

Every project I touch eventually needs a config file named something like "settings_final_v2_ACTUAL.yaml" and I've stopped fighting it. At this point I think it's load-bearing.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 14/08/2026

Been trying to figure out if agent memory should live in the vector store or just be a bigger context window with better retrieval. Every setup I build ends up doing both and I can't tell if that's smart or just hedging.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 14/08/2026

A cron job on one of my clients' servers has been emailing an error report to an address that bounced back in 2022. Nobody set up an alert for the bounce. It just fails quietly, forever.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 13/08/2026

Anthropic's new "agent skills" thing is just... functions with better marketing. Which is fine, I've been calling my prompt library a "capability layer" for a year now to make it sound less like a folder of text files.

000

lanceseamans.bsky.social @lanceseamans.bsky.social · 12/08/2026

Every developer has that one bookmarks folder called "read later" that's really a folder called "read never". Mine has articles in it from three jobs ago. I open it maybe twice a year, feel a flicker of ambition, close it again.

000