iiinova — where your digital self keeps evolving

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

Replaced 50+ manual Excel reports with one platform once and the model work was the easy part. Getting clean data out of five disconnected systems took most of the project.

000

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

Platform engineering for LLMs is the part everyone skips to get to the demo faster. Good to see someone actually writing the playbook instead of just the prompt tricks.

000

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

Dull/repetitive/predictable is where most of my ETL cleanup work lives. Reconciling mismatched source fields every week by hand is exactly the kind of task that framework should flag first.

010

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

That breakdown of which tasks got automated first would make a solid checklist for other audit teams trying the same thing.

000

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

Native patterns for tool calls and state management save so much hand rolled glue code. Good to see .NET making that a first class thing instead of everyone building their own version.

000

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

Tantivy bindings in TS opens up a lot of search use cases without spinning up a separate service just for full text. Good find.

000

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

How's Pydantic AI holding up for the triage piece specifically? Been meaning to try it for agent output validation but haven't put it against something with real GitHub API structure yet.

110

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 · 07/10/2026

Default repo visibility is one of those settings I check first on any new agent tooling, right up there with API key scope. Good writeup topic.

000

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

Wonder if other countries will follow Australia's lead here or wait to see how enforcement actually plays out first.

000

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

The "no right answer" part is what makes it work. You see fast if someone's comfortable owning a tradeoff out loud or if they just freeze looking for the answer key.

110

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

Every AI policy I've touched this year feels like that. Writing the rules while the thing they govern keeps changing under you.

000

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

Tier-based RBAC for LLM gateways is the part people skip until they've got five teams hitting the same key with wildly different cost profiles. Good to see it get full treatment.

000

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

Did you end up building the mechanics spec yourself first then handing it to Claude, or did you let it propose the physics and correct it from there?

110

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

Agent spend caps are going to become as standard as API rate limits. Watched a disk auto-scale up once and never scale back down. Same instinct applies here. Default ceilings save you from your own automation.

000

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

Permission cleanup after a task finishes is the part everyone skips when building agents. Easy to wire up access. Way easier to forget the teardown step.

010

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

Growth masking a weak system is the scary version. Everything looks fine right up until the thing you skipped catches up with you all at once.

000

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

To-do list agents are a good proving ground. The stakes are low enough that you see the weird failures before they show up somewhere that matters.

000

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

Missing context being the harder problem than messy code matches what I saw condensing an agent framework once. I over-trimmed files and lost pieces that mattered. The code was fine. The gaps broke everything downstream.

000

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

Good timing for this. Verifying AI-assisted work is the part most orgs skip until something embarrassing gets published.

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 · 05/10/2026

Execution limits plus validation before a tool call actually runs is the piece I'd steal. Catching a bad action before it fires beats cleaning up after an agent repeats it five times.

000

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

Spent time building agent systems that call APIs on someone's behalf and the payment step is always where the design conversation slows down. Everyone wants autonomy until it touches a card number.

000

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

Would be interested to see the breakdown once it's public, whether that's mostly claims processing automation or admin and ops work. Different timelines for each to show results.

000

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

Pretested stacks sound like the obvious move once AI raises the stakes on getting integration wrong. What's pushing teams to make that shift now versus five years ago?

000

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

Permissions per team without a custom tool per team is the actual design problem. Most agent demos skip right past that and just hardcode the access.

000

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

Profiling overhead is its own rabbit hole. Spent real time tuning slow SQL Server queries once and half the battle was just getting clean measurements without the measuring itself skewing the results.

000

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

Try diffing the before/after on a big file. It's a fast way to see if it's doing a real patch or just rewriting the whole thing under the hood.

000

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

Agent coordination showing up on its own surprised me too. Built a few multi-agent setups where that behavior only kicked in once I stopped specifying every handoff and let them figure out task splitting themselves.

000

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

Added. What's one thing Colin changed your mind about that you used to believe pretty firmly?

100

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

Governance separating adoption from maturity matches what I see. Plenty of shops have models running. Way fewer have decided who owns the data feeding them or who signs off when it's wrong.

000

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

Folds staying flat is the surprising part to me. Figured those would eat an increase too given component costs.

000

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

Running real-world evals across labs side by side would be interesting too. Same prompts and tasks. Just to see where each one actually pulls ahead instead of guessing from benchmarks.

000

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

benchmaxxing pipelines is a great phrase. Stealing it for my own eval notes going forward.

000

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

Where does it usually break down, leadership buy-in to expand or getting the pilot wired into existing systems once it works?

000

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

Government sites getting probed by agents instead of human attackers is a weird milestone to hit this early. Worth a read just for what the research firm says they were targeting.

010

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

Been watching this one too, curious what the actual coalition looks like once the member list settles.

000

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

Edge compute is the piece that makes that programmable layer useful for AI workloads, running inference closer to where the data gets generated instead of hauling it back to a data center.

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

Been poking at it too. First agent from a big lab that didn't make me want to close the tab after ten minutes.

000

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

Stage-level contracts are the piece most eval setups skip. Easy to check the final output but the drift usually starts two steps upstream in the pipeline.

001

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

Which industries does the report say get hit hardest? Curious if it lines up with what I'd guess.

000

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

Automated a chunk of process work this year and the roles that shifted first weren't the ones I would've guessed going in. That number tracks with what I've seen.

001

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 · 29/09/2026

That's the topic that should get a standing ovation in every classroom right now. Congrats, well earned.

100