pflantzer the lab

// the lab of adam pflantzer  ·  est. on live systems

bench://status
forge warm
4 exhibits on record
slot 005 reserved

pflantzer

 a workshop, not a brochure.

What gets built here: harnesses, memory, and proof — the machinery for letting coding agents work on live enterprise systems without breaking them. Four pieces on the bench. All of them real.

index of exhibits status as of 2026-08
  1. 001 Keel a harness for coding agents on live systems private · early
  2. 002 The Platform-Walls Pack 148 lessons in how ServiceNow actually bites audited
  3. 003 The Overnight Verification Loop autonomous scenario testing, dusk to dawn ran in production
  4. 004 Once a public ledger for paying every mistake exactly once live · use it now
  5. 005 — reserved · the bench stays warm
exhibit 001 status: private · early

Keel

A harness for coding agents on live systems.

What it is

Coding agents are excellent in a sandbox and dangerous on a live system — the write that returns 200 but silently did nothing, the change made under the wrong identity, the mistake a team already paid for three weeks ago. None of that is the model being dumb. It's the absence of a harness: the discipline senior engineers apply automatically and agents don't have unless you give it to them.

Keel is that discipline, packaged. An operating doctrine — live system is the source of truth, read-only first, verify every write by read-back, attribute everything, log it. Structural guardrails as policy-as-code: an attribution guard that refuses unattributed writes, protected paths, update-set discipline — hooks that intercept the action before it runs, and ship with a passing self-test. A compounding memory where every burn becomes one crisp, agent-readable fact. And an append-only audit ledger, so nothing about a run is ever a mystery.

Why you can believe it

Keel was not designed on a whiteboard. It was forged doing real, paid enterprise platform work with money on the line — every guard exists because the mistake it blocks actually happened, once, and was made structurally impossible after. The mistakes are in the ledger too. That's what makes the rest of it believable.

exhibit 002 status: audited · three-lens ip scrub

The Platform-Walls Pack

A 148-lesson compounding memory of how ServiceNow actually bites.

What it is

Every platform has walls — the behaviors nobody documents because everybody learns them the expensive way. ACL traps. Update-set capture that quietly puts your changes in the wrong container. Catalog variables that don't behave like fields. Multi-row variable sets. Workspaces. The pack is 148 of those walls, each one earned by hitting it on a live instance — never written speculatively, never padded.

Each lesson is one crisp, agent-readable fact in a fixed anatomy — the symptom you see, the cause underneath, the rule that prevents it — indexed and cross-linked so an agent can reach any wall in one hop. Then the whole pack was scrubbed through a three-lens adversarial IP audit: client names, instance URLs, business vocabulary, and real record data replaced with neutral examples, without altering the platform behavior a single lesson documents.

The idea

An agent that starts already knowing the walls. Not smarter — pre-burned. Every lesson in the pack is a mistake no future instance has to pay for again.

exhibit 003 status: ran on a live instance · unattended

The Overnight Verification Loop

Autonomous scenario testing on a live instance, dusk to dawn.

What it is

UAT is where enterprise delivery goes to stall: weeks of humans clicking through scenarios, screenshotting evidence, losing the thread. The loop replaces the clicking, not the standard of proof. 48 scenarios, run overnight on a live instance, unattended — each one executed, checked, and captured while the humans sleep.

What's waiting in the morning is not a green checkmark. It's a signed UAT evidence workbook: every scenario, what it did, what it verified, what it found. Evidence a stakeholder can hold, not a claim they have to trust.

The discipline underneath

An autonomous run is only as trustworthy as its worst moment. When something breaks mid-run, the loop's rule is absolute: the incident is logged and fixed on the record — never hidden, never quietly retried out of existence. The workbook includes the stumble and the recovery, because a test run that can't admit failure isn't testing anything.

exhibit 004 status: live · free · open to everyone

Once

A public instrument: pay for every mistake exactly once.

What it is

Every engineer carries a private list of scars — the deploy that lied, the outage that happened twice, the fix that left with the person who knew it. Once is a small working tool for writing those scars down in the lab's lesson anatomy — symptom, cause, rule — and it's open to anyone. Your ledger lives entirely in your own browser: no server, no account, no tracker. One click exports it as agent-readable markdown you can paste straight into a CLAUDE.md, an AGENTS.md, a runbook, or a new teammate's onboarding doc.

Why it exists

The world would be a better place if every hard-won lesson were written down once, in a form both people and machines could learn from — instead of being paid for again by the next person to hit the same wall. The ticker at the top of this page has said pay for a mistake once since the lab opened. This is that line, turned into an instrument and handed to everyone.

open the instrument → /once

// the hands behind the bench

Everything above exists because a client needed it to.

I'm Adam Pflantzer — an independent ServiceNow architect. The lab is where the instruments get built; the practice is where they earn their keep, on live instances, with money on the line. If you want this discipline pointed at your platform, that conversation starts next door.

work with me → adam.pflantzer.com email the lab