Karl Lowenbjer
Jul 24, 2026 - Sep 7, 2026
4
Decisions
0
Active Todos
2
Patterns
Categories
Decisions (4)
Extend Karl Lowenbjer under Wallace and land Atlas and Motherbrain in Engineering
After spending the prior week reading the Motherbrain and Atlas code Karl Lowenbjer built under Tabatha, Peter concluded it was quite good - really well architected and thought through. Chris Baek did not want to keep Karl on the Operations side. In the Thursday 1:1 Peter pitched Wallace directly on taking both the contractor and the applications, telling him he is pretty sure Wallace ends up being the one asked to maintain things like Motherbrain and Atlas anyway. Wallace agreed the same call, having independently run a compliance and security read of the code and landed in the same place. On Saturday Peter DMed Karl himself, told him his contract ends in about a week and a half, and asked whether he wanted to let it end or extend it under Peter - then asked Sarah for 30 minutes with Karl on Tuesday or Wednesday to work it out.
Judge the Atlas code on its engineering merits alone - keep it, port it off the laptop, and stop building skills with direct data access
After a deep dive through the revops-motherbrain repo and a walkthrough with Karl Lowenbjer, Tabatha Wilmot and Chris Baek, Peter delivered a written verdict on 8/31. He opened by refusing the question he was not asked: I am making ZERO statements about whether what this code was written to do is something we need at CIQ - that is a discussion for others. On the code itself: keep the bulk of it. API, frontend, warehouse and agents are independently separable with no cross dependencies. Data access is SQL direct rather than through hallucination-prone skill paths, read-only is enforced at the right layers, the MCP server sits above those layers, test-to-code ratio is roughly 2:3 and the documentation has not drifted. Fixes named: pull the business logic out of the code into a human-readable rule set (he called it the most valuable part of the project), delete the dead Trust Anchor, remove a write token that should be read-only, de-duplicate the mql LLM screen defined in two codebases. The real problem is what runs on the contractor laptop - Claude skills against the claude.ai MCP connector with direct HubSpot write scope, running without asking permission, bypassing the pipeline-approval and closed-edit rules HubSpot enforces on humans. Verdict on the open PRs: approve the move to a consolidated server, do NOT proceed with the interim laptop-resident lead router, fix it properly instead. And a standing instruction: the process of building out skills with direct access to data should stop immediately. In the 8/31 sync with Bjorn and Chris Baek he also confirmed he could pick up and make good use of the code if the contractor who wrote it did not stay - separating the keep-the-code question from the keep-the-person question.
Scope the Atlas review to how it works, not whether it is worth it - and restate the rule that Claude must call a script, never an MCP directly
In the Motherbrain / Atlas sync with Tabatha Wilmot, Karl Lowenbjer and Chris Baek, Peter set the terms of his own code review. He told them outright that he would not wade into whether Atlas is valuable - that is a Tabatha-and-sales question - and that what he wants to answer is how it works, how it is architected, how it stores and persists data, how it does security protections, what should run on a laptop versus centralized, and what the scheduler should be. He asked for the scripts and skills that still live only on Tabatha laptop to be pushed into the repo first so he can read both ends of the system, asked for half-sentence direct answers to Marlon open questions rather than more narrative, and committed to a four-to-five-hour deep dive by Saturday evening with a follow-up the middle of next week. When Tabatha described reps closing deals wrongly through the HubSpot MCP, Peter named the cause - Claude writing code at runtime against a system of record - and restated the standing rule that Claude should never call an MCP directly, it should call a script that structurally cannot do that.
Require all agent/LLM access to systems of record to go through inspectable static code, never direct LLM access
Peter opened the 7/24 Jira Automation call with Karl Lowenbjer by setting a condition before taking Karls agenda. The rule: no LLM gets direct access to a system of record. In every instance the LLM writes code that accesses the system, and that code is inspectable by a human or by another LLM. He stated he wants this true for everything Karls team builds, and that as long as it is true he is comfortable. He explicitly accepted that the filters themselves may be wrong - we might make a mistake in terms of what we pull in and out and that is fine, I do not mind mistakes - but the mechanism is non-negotiable. Karl confirmed Atlas already works this way. Peter then closed the topic without further discussion.
Related Patterns (2)
Redesign Conditions Over Policing Symptoms
When a direct report names a vulnerability and proposes surveillance-style verification mechanisms (breathalyzers, daily check-ins, monitoring rituals), Peter accepts the disclosure but pushes back on the surveillance model. Treats the verification proposal as a signal that the underlying environment needs redesign — and offers to change the conditions that produced the vulnerability rather than instrument the symptom. Costs Peter optionality (e.g., committing to broker work-hour expectations directly with the partner/spouse) — the asymmetry signals genuine retention vs transactional.
Constrain the Mechanism, Not the Access
When AI or automation raises a governance concern, Peter does not restrict who or what may reach the data. He constrains HOW the reach happens, so the pathway becomes auditable and deterministic. The failure mode he designs against is non-determinism, not leakage - a wrong filter in inspectable static code is a bug you fix once, while a model that sometimes pulls the wrong field is an unbounded, unauditable liability. He will therefore accept imperfect filtering and eventual leakage as the price of adoption speed, but will not accept a non-inspectable pathway.